The Apache HBase™ Reference Guide

Revision History
Revision 2014-07-08T15:28


This is the official reference guide of Apache HBase™, a distributed, versioned, column-oriented database built on top of Apache Hadoop™ and Apache ZooKeeper™.

Table of Contents

1. Getting Started
1.1. Introduction
1.2. Quick Start
2. Apache HBase Configuration
2.1. Basic Prerequisites
2.2. HBase run modes: Standalone and Distributed
2.3. Configuration Files
2.4. Example Configurations
2.5. The Important Configurations
3. Upgrading
3.1. HBase version numbers
3.2. Upgrading from 0.94.x to 0.96.x
3.3. Upgrading from 0.92.x to 0.94.x
3.4. Upgrading from 0.90.x to 0.92.x
3.5. Upgrading to HBase 0.90.x from 0.20.x or 0.89.x
4. The Apache HBase Shell
4.1. Scripting
4.2. Shell Tricks
5. Data Model
5.1. Conceptual View
5.2. Physical View
5.3. Namespace
5.4. Table
5.5. Row
5.6. Column Family
5.7. Cells
5.8. Data Model Operations
5.9. Versions
5.10. Sort Order
5.11. Column Metadata
5.12. Joins
5.13. ACID
6. HBase and Schema Design
6.1. Schema Creation
6.2. On the number of column families
6.3. Rowkey Design
6.4. Number of Versions
6.5. Supported Datatypes
6.6. Joins
6.7. Time To Live (TTL)
6.8. Keeping Deleted Cells
6.9. Secondary Indexes and Alternate Query Paths
6.10. Constraints
6.11. Schema Design Case Studies
6.12. Operational and Performance Configuration Options
7. HBase and MapReduce
7.1. Map-Task Splitting
7.2. HBase MapReduce Examples
7.3. Accessing Other HBase Tables in a MapReduce Job
7.4. Speculative Execution
8. Secure Apache HBase
8.1. Secure Client Access to Apache HBase
8.2. Access Control
8.3. Secure Bulk Load
9. Architecture
9.1. Overview
9.2. Catalog Tables
9.3. Client
9.4. Client Request Filters
9.5. Master
9.6. RegionServer
9.7. Regions
9.8. Bulk Loading
9.9. HDFS
10. Apache HBase External APIs
10.1. Non-Java Languages Talking to the JVM
10.2. REST
10.3. Thrift
10.4. C/C++ Apache HBase Client
11. Apache HBase Coprocessors
12. Apache HBase Performance Tuning
12.1. Operating System
12.2. Network
12.3. Java
12.4. HBase Configurations
12.5. ZooKeeper
12.6. Schema Design
12.7. HBase General Patterns
12.8. Writing to HBase
12.9. Reading from HBase
12.10. Deleting from HBase
12.11. HDFS
12.12. Amazon EC2
12.13. Collocating HBase and MapReduce
12.14. Case Studies
13. Troubleshooting and Debugging Apache HBase
13.1. General Guidelines
13.2. Logs
13.3. Resources
13.4. Tools
13.5. Client
13.6. MapReduce
13.7. NameNode
13.8. Network
13.9. RegionServer
13.10. Master
13.11. ZooKeeper
13.12. Amazon EC2
13.13. HBase and Hadoop version issues
13.14. Running unit or integration tests
13.15. Case Studies
14. Apache HBase Case Studies
14.1. Overview
14.2. Schema Design
14.3. Performance/Troubleshooting
15. Apache HBase Operational Management
15.1. HBase Tools and Utilities
15.2. Region Management
15.3. Node Management
15.4. HBase Metrics
15.5. HBase Monitoring
15.6. Cluster Replication
15.7. HBase Backup
15.8. HBase Snapshots
15.9. Capacity Planning and Region Sizing
15.10. Table Rename
16. Building and Developing Apache HBase
16.1. Apache HBase Repositories
16.2. IDEs
16.3. Building Apache HBase
16.4. Releasing Apache HBase
16.5. Generating the HBase Reference Guide
16.6. Updating
16.7. Tests
16.8. Maven Build Commands
16.9. Getting Involved
16.10. Developing
16.11. Submitting Patches
17. ZooKeeper
17.1. Using existing ZooKeeper ensemble
17.2. SASL Authentication with ZooKeeper
18. Apache HBase Coprocessors
19. Community
19.1. Decisions
19.2. Community Roles
B. hbck In Depth
B.1. Running hbck to identify inconsistencies
B.2. Inconsistencies
B.3. Localized repairs
B.4. Region Overlap Repairs
C. Compression In HBase
C.1. CompressionTest Tool
C.2. hbase.regionserver.codecs
C.3. LZO
C.6. Changing Compression Schemes
D. YCSB: The Yahoo! Cloud Serving Benchmark and HBase
E. HFile format version 2
E.1. Motivation
E.2. HFile format version 1 overview
E.3. HBase file format with inline blocks (version 2)
F. Other Information About HBase
F.1. HBase Videos
F.2. HBase Presentations (Slides)
F.3. HBase Papers
F.4. HBase Sites
F.5. HBase Books
F.6. Hadoop Books
G. HBase History
H. HBase and the Apache Software Foundation
H.1. ASF Development Process
H.2. ASF Board Reporting
I. Enabling Dapper-like Tracing in HBase
I.1. SpanReceivers
I.2. Client Modifications
J. 0.95 RPC Specification
J.1. Goals
J.3. RPC
J.4. Notes

List of Tables

2.1. Hadoop version support matrix
5.1. Table webtable
5.2. ColumnFamily anchor
5.3. ColumnFamily contents
8.1. Operation To Permission Mapping


This is the official reference guide for the HBase version it ships with. Herein you will find either the definitive documentation on an HBase topic as of its standing when the referenced HBase version shipped, or it will point to the location in javadoc, JIRA or wiki where the pertinent information can be found.

This reference guide is a work in progress. The source for this guide can be found at src/main/docbkx in a checkout of the hbase project. This reference guide is marked up using DocBook from which the the finished guide is generated as part of the 'site' build target: i.e. run

$mvn site

to generate this documentation. Feel free to add content by adding a patch to an issue up in the HBase JIRA.


If this is your first foray into the wonderful world of Distributed Computing, then you are in for some interesting times. First off, distributed systems are hard; making a distributed system hum requires a disparate skillset that spans systems (hardware and software) and networking. Your cluster' operation can hiccup because of any of a myriad set of reasons from bugs in HBase itself through misconfigurations -- misconfiguration of HBase but also operating system misconfigurations -- through to hardware problems whether it be a bug in your network card drivers or an underprovisioned RAM bus (to mention two recent examples of hardware issues that manifested as "HBase is slow"). You will also need to do a recalibration if up to this your computing has been bound to a single box. Here is one good starting point: Fallacies of Distributed Computing.

Chapter 1. Getting Started

1.1. Introduction

Section 1.2, “Quick Start” will get you up and running on a single-node instance of HBase using the local filesystem.

1.2. Quick Start

This guide describes setup of a standalone HBase instance that uses the local filesystem. It leads you through creating a table, inserting rows via the HBase shell, and then cleaning up and shutting down your standalone HBase instance. The below exercise should take no more than ten minutes (not including download time).

Local Filesystem and Durability

Using HBase with a LocalFileSystem does not currently guarantee durability. See and its associated issues for more details.

Loopback IP

HBase expects the loopback IP address to be Ubuntu and some other distributions, for example, will default to and this will cause problems for you [1].

/etc/hosts should look something like this:

   ubuntu.ubuntu-domain ubuntu

Fixed in 0.96.0+

As of hbase-0.96.0+, this should no longer be an issue; it should work w/o special modification of /etc/hosts

1.2.1. Download and unpack the latest stable release.

Choose a download site from this list of Apache Download Mirrors. Click on the suggested top link. This will take you to a mirror of HBase Releases. Click on the folder named stable and then download the file that ends in .tar.gz to your local filesystem; e.g. hbase-0.94.2.tar.gz.

Decompress and untar your download and then change into the unpacked directory.

$ tar xfz hbase-
$ cd hbase-

At this point, you are ready to start HBase. But before starting it, edit conf/hbase-site.xml, the file you write your site-specific configurations into. Set hbase.rootdir, the directory HBase writes data to, and, the director ZooKeeper writes its data too:

<?xml version="1.0"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>

Replace DIRECTORY in the above with the path to the directory you would have HBase and ZooKeeper write their data. By default, hbase.rootdir is set to /tmp/hbase-${} and similarly so for the default ZooKeeper data location which means you'll lose all your data whenever your server reboots unless you change it (Most operating systems clear /tmp on restart).

1.2.2. Start HBase

Now start HBase:

$ ./bin/
starting Master, logging to logs/

You should now have a running standalone HBase instance. In standalone mode, HBase runs all daemons in the the one JVM; i.e. both the HBase and ZooKeeper daemons. HBase logs can be found in the logs subdirectory. Check them out especially if it seems HBase had trouble starting.

Is java installed?

All of the above presumes a 1.6 version of Oracle java is installed on your machine and available on your path (See Section 2.1.1, “Java”); i.e. when you type java, you see output that describes the options the java program takes (HBase requires java 6). If this is not the case, HBase will not start. Install java, edit conf/, uncommenting the JAVA_HOME line pointing it to your java install, then, retry the steps above.

1.2.3. Shell Exercises

Connect to your running HBase via the shell.

$ ./bin/hbase shell
HBase Shell; enter 'help<RETURN>' for list of supported commands.
Type "exit<RETURN>" to leave the HBase Shell
Version: 0.90.0, r1001068, Fri Sep 24 13:55:42 PDT 2010


Type help and then <RETURN> to see a listing of shell commands and options. Browse at least the paragraphs at the end of the help emission for the gist of how variables and command arguments are entered into the HBase shell; in particular note how table names, rows, and columns, etc., must be quoted.

Create a table named test with a single column family named cf. Verify its creation by listing all tables and then insert some values.

hbase(main):003:0> create 'test', 'cf'
0 row(s) in 1.2200 seconds
hbase(main):003:0> list 'test'
1 row(s) in 0.0550 seconds
hbase(main):004:0> put 'test', 'row1', 'cf:a', 'value1'
0 row(s) in 0.0560 seconds
hbase(main):005:0> put 'test', 'row2', 'cf:b', 'value2'
0 row(s) in 0.0370 seconds
hbase(main):006:0> put 'test', 'row3', 'cf:c', 'value3'
0 row(s) in 0.0450 seconds

Above we inserted 3 values, one at a time. The first insert is at row1, column cf:a with a value of value1. Columns in HBase are comprised of a column family prefix -- cf in this example -- followed by a colon and then a column qualifier suffix (a in this case).

Verify the data insert by running a scan of the table as follows

hbase(main):007:0> scan 'test'
row1       column=cf:a, timestamp=1288380727188, value=value1
row2       column=cf:b, timestamp=1288380738440, value=value2
row3       column=cf:c, timestamp=1288380747365, value=value3
3 row(s) in 0.0590 seconds

Get a single row

hbase(main):008:0> get 'test', 'row1'
cf:a        timestamp=1288380727188, value=value1
1 row(s) in 0.0400 seconds

Now, disable and drop your table. This will clean up all done above.

hbase(main):012:0> disable 'test'
0 row(s) in 1.0930 seconds
hbase(main):013:0> drop 'test'
0 row(s) in 0.0770 seconds 

Exit the shell by typing exit.

hbase(main):014:0> exit

1.2.4. Stopping HBase

Stop your hbase instance by running the stop script.

$ ./bin/
stopping hbase...............

1.2.5. Where to go next

The above described standalone setup is good for testing and experiments only. In the next chapter, Chapter 2, Apache HBase Configuration, we'll go into depth on the different HBase run modes, system requirements running HBase, and critical configurations setting up a distributed HBase deploy.

Chapter 2. Apache HBase Configuration

This chapter is the Not-So-Quick start guide to Apache HBase configuration. It goes over system requirements, Hadoop setup, the different Apache HBase run modes, and the various configurations in HBase. Please read this chapter carefully. At a mimimum ensure that all Section 2.1, “Basic Prerequisites” have been satisfied. Failure to do so will cause you (and us) grief debugging strange errors and/or data loss.

Apache HBase uses the same configuration system as Apache Hadoop. To configure a deploy, edit a file of environment variables in conf/ -- this configuration is used mostly by the launcher shell scripts getting the cluster off the ground -- and then add configuration to an XML file to do things like override HBase defaults, tell HBase what Filesystem to use, and the location of the ZooKeeper ensemble [2] .

When running in distributed mode, after you make an edit to an HBase configuration, make sure you copy the content of the conf directory to all nodes of the cluster. HBase will not do this for you. Use rsync.

2.1. Basic Prerequisites

This section lists required services and some required system configuration.

2.1.1. Java

Just like Hadoop, HBase requires at least Java 6 from Oracle. Java 7 should work and can even be faster than Java 6, but almost all testing has been done on the latter at this point.

2.1.2. Operating System ssh

ssh must be installed and sshd must be running to use Hadoop's scripts to manage remote Hadoop and HBase daemons. You must be able to ssh to all nodes, including your local node, using passwordless login (Google "ssh passwordless login"). If on mac osx, see the section, SSH: Setting up Remote Desktop and Enabling Self-Login on the hadoop wiki. DNS

HBase uses the local hostname to self-report its IP address. Both forward and reverse DNS resolving must work in versions of HBase previous to 0.92.0 [3].

If your machine has multiple interfaces, HBase will use the interface that the primary hostname resolves to.

If this is insufficient, you can set hbase.regionserver.dns.interface to indicate the primary interface. This only works if your cluster configuration is consistent and every host has the same network interface configuration.

Another alternative is setting hbase.regionserver.dns.nameserver to choose a different nameserver than the system wide default. Loopback IP

HBase expects the loopback IP address to be See Section, “Loopback IP” NTP

The clocks on cluster members should be in basic alignments. Some skew is tolerable but wild skew could generate odd behaviors. Run NTP on your cluster, or an equivalent.

If you are having problems querying data, or "weird" cluster operations, check system time!  ulimit and nproc

Apache HBase is a database. It uses a lot of files all at the same time. The default ulimit -n -- i.e. user file limit -- of 1024 on most *nix systems is insufficient (On mac os x its 256). Any significant amount of loading will lead you to Section, “ many open files)”. You may also notice errors such as...

      2010-04-06 03:04:37,542 INFO org.apache.hadoop.hdfs.DFSClient: Exception increateBlockOutputStream
      2010-04-06 03:04:37,542 INFO org.apache.hadoop.hdfs.DFSClient: Abandoning block blk_-6935524980745310745_1391901

Do yourself a favor and change the upper bound on the number of file descriptors. Set it to north of 10k. The math runs roughly as follows: per ColumnFamily there is at least one StoreFile and possibly up to 5 or 6 if the region is under load. Multiply the average number of StoreFiles per ColumnFamily times the number of regions per RegionServer. For example, assuming that a schema had 3 ColumnFamilies per region with an average of 3 StoreFiles per ColumnFamily, and there are 100 regions per RegionServer, the JVM will open 3 * 3 * 100 = 900 file descriptors (not counting open jar files, config files, etc.)

You should also up the hbase users' nproc setting; under load, a low-nproc setting could manifest as OutOfMemoryError [4] [5].

To be clear, upping the file descriptors and nproc for the user who is running the HBase process is an operating system configuration, not an HBase configuration. Also, a common mistake is that administrators will up the file descriptors for a particular user but for whatever reason, HBase will be running as some one else. HBase prints in its logs as the first line the ulimit its seeing. Ensure its correct. [6] ulimit on Ubuntu

If you are on Ubuntu you will need to make the following changes:

In the file /etc/security/limits.conf add a line like:

hadoop  -       nofile  32768

Replace hadoop with whatever user is running Hadoop and HBase. If you have separate users, you will need 2 entries, one for each user. In the same file set nproc hard and soft limits. For example:

hadoop soft/hard nproc 32000


In the file /etc/pam.d/common-session add as the last line in the file:

session required

Otherwise the changes in /etc/security/limits.conf won't be applied.

Don't forget to log out and back in again for the changes to take effect! Windows

Apache HBase has been little tested running on Windows. Running a production install of HBase on top of Windows is not recommended.

If you are running HBase on Windows, you must install Cygwin to have a *nix-like environment for the shell scripts. The full details are explained in the Windows Installation guide. Also search our user mailing list to pick up latest fixes figured by Windows users.

2.1.3. Hadoop

Selecting a Hadoop version is critical for your HBase deployment. Below table shows some information about what versions of Hadoop are supported by various HBase versions. Based on the version of HBase, you should select the most appropriate version of Hadoop. We are not in the Hadoop distro selection business. You can use Hadoop distributions from Apache, or learn about vendor distributions of Hadoop at

Hadoop 2.x is better than Hadoop 1.x

Hadoop 2.x is faster, with more features such as short-circuit reads which will help improve your HBase random read profile as well important bug fixes that will improve your overall HBase experience. You should run Hadoop 2. rather than Hadoop 1. if you can.

Table 2.1. Hadoop version support matrix

Hadoop-0.22.x SXX
Hadoop-1.0.0-1.0.2[a] SSX
Hadoop-1.1.x NTSS
Hadoop-0.23.x XSNT
Hadoop-2.0.x-alpha XNTX
Hadoop-2.1.0-beta XNTS
Hadoop-2.x XNTS

[a] HBase requires hadoop 1.0.3 at a minimum; there is an issue where we cannot find KerberosUtil compiling against earlier versions of Hadoop.


S = supported and tested,
X = not supported,
NT = it should run, but not tested enough.

Because HBase depends on Hadoop, it bundles an instance of the Hadoop jar under its lib directory. The bundled jar is ONLY for use in standalone mode. In distributed mode, it is critical that the version of Hadoop that is out on your cluster match what is under HBase. Replace the hadoop jar found in the HBase lib directory with the hadoop jar you are running on your cluster to avoid version mismatch issues. Make sure you replace the jar in HBase everywhere on your cluster. Hadoop version mismatch issues have various manifestations but often all looks like its hung up. Apache HBase 0.92 and 0.94

HBase 0.92 and 0.94 versions can work with Hadoop versions, 0.20.205, 0.22.x, 1.0.x, and 1.1.x. HBase-0.94 can additionally work with Hadoop-0.23.x and 2.x, but you may have to recompile the code using the specific maven profile (see top level pom.xml) Apache HBase 0.96

Apache HBase 0.96.0 requires Apache Hadoop 1.x at a minimum, and it can run equally well on hadoop-2.0. As of Apache HBase 0.96.x, Apache Hadoop 1.0.x at least is required. We will no longer run properly on older Hadoops such as 0.20.205 or branch-0.20-append. Do not move to Apache HBase 0.96.x if you cannot upgrade your Hadoop[7]. Hadoop versions 0.20.x - 1.x

HBase will lose data unless it is running on an HDFS that has a durable sync implementation. DO NOT use Hadoop 0.20.2, Hadoop, and Hadoop which DO NOT have this attribute. Currently only Hadoop versions 0.20.205.x or any release in excess of this version -- this includes hadoop-1.0.0 -- have a working, durable sync [8]. Sync has to be explicitly enabled by setting equal to true on both the client side -- in hbase-site.xml -- and on the serverside in hdfs-site.xml (The sync facility HBase needs is a subset of the append code path).


You will have to restart your cluster after making this edit. Ignore the chicken-little comment you'll find in the hdfs-default.xml in the description for the configuration. Apache HBase on Secure Hadoop

Apache HBase will run on any Hadoop 0.20.x that incorporates Hadoop security features as long as you do as suggested above and replace the Hadoop jar that ships with HBase with the secure version. If you want to read more about how to setup Secure HBase, see Section 8.1, “Secure Client Access to Apache HBase”. dfs.datanode.max.xcievers

An Hadoop HDFS datanode has an upper bound on the number of files that it will serve at any one time. The upper bound parameter is called xcievers (yes, this is misspelled). Again, before doing any loading, make sure you have configured Hadoop's conf/hdfs-site.xml setting the xceivers value to at least the following:


Be sure to restart your HDFS after making the above configuration.

Not having this configuration in place makes for strange looking failures. Eventually you'll see a complain in the datanode logs complaining about the xcievers exceeded, but on the run up to this one manifestation is complaint about missing blocks. For example: 10/12/08 20:10:31 INFO hdfs.DFSClient: Could not obtain block blk_XXXXXXXXXXXXXXXXXXXXXX_YYYYYYYY from any node: No live nodes contain current block. Will get new block locations from namenode and retry... [9]

See also Section 14.3.4, “Case Study #4 (xcievers Config)”

2.2. HBase run modes: Standalone and Distributed

HBase has two run modes: Section 2.2.1, “Standalone HBase” and Section 2.2.2, “Distributed”. Out of the box, HBase runs in standalone mode. Whatever your mode, you will need to configure HBase by editing files in the HBase conf directory. At a minimum, you must edit conf/ to tell HBase which java to use. In this file you set HBase environment variables such as the heapsize and other options for the JVM, the preferred location for log files, etc. Set JAVA_HOME to point at the root of your java install.

2.2.1. Standalone HBase

This is the default mode. Standalone mode is what is described in the Section 1.2, “Quick Start” section. In standalone mode, HBase does not use HDFS -- it uses the local filesystem instead -- and it runs all HBase daemons and a local ZooKeeper all up in the same JVM. Zookeeper binds to a well known port so clients may talk to HBase.

2.2.2. Distributed

Distributed mode can be subdivided into distributed but all daemons run on a single node -- a.k.a pseudo-distributed-- and fully-distributed where the daemons are spread across all nodes in the cluster [10].

Pseudo-distributed mode can run against the local filesystem or it can run against an instance of the Hadoop Distributed File System (HDFS). Fully-distributed mode can ONLY run on HDFS. See the Hadoop requirements and instructions for how to set up HDFS.

Below we describe the different distributed setups. Starting, verification and exploration of your install, whether a pseudo-distributed or fully-distributed configuration is described in a section that follows, Section 2.2.3, “Running and Confirming Your Installation”. The same verification script applies to both deploy types. Pseudo-distributed

A pseudo-distributed mode is simply a fully-distributed mode run on a single host. Use this configuration testing and prototyping on HBase. Do not use this configuration for production nor for evaluating HBase performance.

First, if you want to run on HDFS rather than on the local filesystem, setup your HDFS. You can set up HDFS also in pseudo-distributed mode. Ensure you have a working HDFS before proceeding.

Next, configure HBase. Edit conf/hbase-site.xml. This is the file into which you add local customizations and overrides. At a minimum, you must tell HBase to run in (pseudo-)distributed mode rather than in default standalone mode. To do this, set the hbase.cluster.distributed property to true (Its default is false). The absolute bare-minimum hbase-site.xml is therefore as follows:


With this configuration, HBase will start up an HBase Master process, a ZooKeeper server, and a RegionServer process running against the local filesystem writing to wherever your operating system stores temporary files into a directory named hbase-YOUR_USER_NAME.

Such a setup, using the local filesystem and writing to the operating systems's temporary directory is an ephemeral setup; the Hadoop local filesystem -- which is what HBase uses when it is writing the local filesytem does not support sync so unless the system is shutdown properly, the data will be lost. Writing to the operating system's temporary directory can also make for data loss when the machine is restarted as this directory is usually cleared on reboot. For a more permanent setup, see the next example where we make use of an instance of HDFS; HBase data will be written to the Hadoop distributed filesystem rather than to the local filesystem's tmp directory.

In this conf/hbase-site.xml example, the hbase.rootdir property points to the local HDFS instance homed on the node

Let HBase create ${hbase.rootdir}

Let HBase create the hbase.rootdir directory. If you don't, you'll get warning saying HBase needs a migration run because the directory is missing files expected by HBase (it'll create them if you let it).


Now skip to Section 2.2.3, “Running and Confirming Your Installation” for how to start and verify your pseudo-distributed install. [11] Pseudo-distributed Extras Startup

To start up the initial HBase cluster...

% bin/

To start up an extra backup master(s) on the same server run...

% bin/ start 1

... the '1' means use ports 60001 & 60011, and this backup master's logfile will be at logs/hbase-${USER}-1-master-${HOSTNAME}.log.

To startup multiple backup masters run...

% bin/ start 2 3

You can start up to 9 backup masters (10 total).

To start up more regionservers...

% bin/ start 1

where '1' means use ports 60201 & 60301 and its logfile will be at logs/hbase-${USER}-1-regionserver-${HOSTNAME}.log.

To add 4 more regionservers in addition to the one you just started by running...

% bin/ start 2 3 4 5

This supports up to 99 extra regionservers (100 total). Stop

Assuming you want to stop master backup # 1, run...

% cat /tmp/hbase-${USER} |xargs kill -9

Note that bin/ stop 1 will try to stop the cluster along with the master.

To stop an individual regionserver, run...

% bin/ stop 1

For running a fully-distributed operation on more than one host, make the following configurations. In hbase-site.xml, add the property hbase.cluster.distributed and set it to true and point the HBase hbase.rootdir at the appropriate HDFS NameNode and location in HDFS where you would like HBase to write data. For example, if you namenode were running at on port 8020 and you wanted to home your HBase in HDFS at /hbase, make the following configuration.

    <description>The directory shared by RegionServers.
    <description>The mode the cluster will be in. Possible values are
      false: standalone and pseudo-distributed setups with managed Zookeeper
      true: fully-distributed with unmanaged Zookeeper Quorum (see
</configuration> regionservers

In addition, a fully-distributed mode requires that you modify conf/regionservers. The Section, “regionservers file lists all hosts that you would have running HRegionServers, one host per line (This file in HBase is like the Hadoop slaves file). All servers listed in this file will be started and stopped when HBase cluster start or stop is run. ZooKeeper and HBase

See section Chapter 17, ZooKeeper for ZooKeeper setup for HBase. HDFS Client Configuration

Of note, if you have made HDFS client configuration on your Hadoop cluster -- i.e. configuration you want HDFS clients to use as opposed to server-side configurations -- HBase will not see this configuration unless you do one of the following:

  • Add a pointer to your HADOOP_CONF_DIR to the HBASE_CLASSPATH environment variable in

  • Add a copy of hdfs-site.xml (or hadoop-site.xml) or, better, symlinks, under ${HBASE_HOME}/conf, or

  • if only a small set of HDFS client configurations, add them to hbase-site.xml.

An example of such an HDFS client configuration is dfs.replication. If for example, you want to run with a replication factor of 5, hbase will create files with the default of 3 unless you do the above to make the configuration available to HBase.

2.2.3. Running and Confirming Your Installation

Make sure HDFS is running first. Start and stop the Hadoop HDFS daemons by running bin/ over in the HADOOP_HOME directory. You can ensure it started properly by testing the put and get of files into the Hadoop filesystem. HBase does not normally use the mapreduce daemons. These do not need to be started.

If you are managing your own ZooKeeper, start it and confirm its running else, HBase will start up ZooKeeper for you as part of its start process.

Start HBase with the following command:

Run the above from the HBASE_HOME directory.

You should now have a running HBase instance. HBase logs can be found in the logs subdirectory. Check them out especially if HBase had trouble starting.

HBase also puts up a UI listing vital attributes. By default its deployed on the Master host at port 60010 (HBase RegionServers listen on port 60020 by default and put up an informational http server at 60030). If the Master were running on a host named on the default port, to see the Master's homepage you'd point your browser at

Once HBase has started, see the Section 1.2.3, “Shell Exercises” for how to create tables, add data, scan your insertions, and finally disable and drop your tables.

To stop HBase after exiting the HBase shell enter

$ ./bin/
stopping hbase...............

Shutdown can take a moment to complete. It can take longer if your cluster is comprised of many machines. If you are running a distributed operation, be sure to wait until HBase has shut down completely before stopping the Hadoop daemons.

2.3. Configuration Files

2.3.1. hbase-site.xml and hbase-default.xml

Just as in Hadoop where you add site-specific HDFS configuration to the hdfs-site.xml file, for HBase, site specific customizations go into the file conf/hbase-site.xml. For the list of configurable properties, see Section, “HBase Default Configuration” below or view the raw hbase-default.xml source file in the HBase source code at src/main/resources.

Not all configuration options make it out to hbase-default.xml. Configuration that it is thought rare anyone would change can exist only in code; the only way to turn up such configurations is via a reading of the source code itself.

Currently, changes here will require a cluster restart for HBase to notice the change. HBase Default Configuration

HBase Default Configuration

The documentation below is generated using the default hbase configuration file, hbase-default.xml, as source.


Temporary directory on the local filesystem. Change this setting to point to a location more permanent than '/tmp', the usual resolve for, as the '/tmp' directory is cleared on machine restart.

Default: ${}/hbase-${}


The directory shared by region servers and into which HBase persists. The URL should be 'fully-qualified' to include the filesystem scheme. For example, to specify the HDFS directory '/hbase' where the HDFS instance's namenode is running at on port 9000, set this value to: hdfs:// By default, we write to whatever ${hbase.tmp.dir} is set too -- usually /tmp -- so change this configuration or else all data will be lost on machine restart.

Default: ${hbase.tmp.dir}/hbase


The mode the cluster will be in. Possible values are false for standalone mode and true for distributed mode. If false, startup will run all HBase and ZooKeeper daemons together in the one JVM.

Default: false


Comma separated list of servers in the ZooKeeper ensemble (This config. should have been named hbase.zookeeper.ensemble). For example, ",,". By default this is set to localhost for local and pseudo-distributed modes of operation. For a fully-distributed setup, this should be set to a full list of ZooKeeper ensemble servers. If HBASE_MANAGES_ZK is set in this is the list of servers which hbase will start/stop ZooKeeper on as part of cluster start/stop. Client-side, we will take this list of ensemble members and put it together with the hbase.zookeeper.clientPort config. and pass it into zookeeper constructor as the connectString parameter.

Default: localhost


Directory on the local filesystem to be used as a local storage.

Default: ${hbase.tmp.dir}/local/


The port the HBase Master should bind to.

Default: 60000

The port for the HBase Master web UI. Set to -1 if you do not want a UI instance run.

Default: 60010

The bind address for the HBase Master web UI



A comma-separated list of LogCleanerDelegate invoked by the LogsCleaner service. These WAL/HLog cleaners are called in order, so put the HLog cleaner that prunes the most HLog files in front. To implement your own LogCleanerDelegate, just put it in HBase's classpath and add the fully qualified class name here. Always add the above default log cleaners in the list.

Default: org.apache.hadoop.hbase.master.cleaner.TimeToLiveLogCleaner


Maximum time a HLog can stay in the .oldlogdir directory, after which it will be cleaned by a Master thread.

Default: 600000


A comma-separated list of HFileCleanerDelegate invoked by the HFileCleaner service. These HFiles cleaners are called in order, so put the cleaner that prunes the most files in front. To implement your own HFileCleanerDelegate, just put it in HBase's classpath and add the fully qualified class name here. Always add the above default log cleaners in the list as they will be overwritten in hbase-site.xml.

Default: org.apache.hadoop.hbase.master.cleaner.TimeToLiveHFileCleaner


Timeout value for the Catalog Janitor from the master to META.

Default: 600000

If abort immediately for the expired master without trying to recover its zk session.

Default: false


The name of the Network Interface from which a master should report its IP address.

Default: default


The host name or IP address of the name server (DNS) which a master should use to determine the host name used for communication and display purposes.

Default: default


The port the HBase RegionServer binds to.

Default: 60020

The port for the HBase RegionServer web UI Set to -1 if you do not want the RegionServer UI to run.

Default: 60030

The address for the HBase RegionServer web UI


Whether or not the Master or RegionServer UI should search for a port to bind to. Enables automatic port search if is already in use. Useful for testing, turned off by default.

Default: false


Count of RPC Listener instances spun up on RegionServers. Same property is used by the Master for count of master handlers.

Default: 30


Interval between messages from the RegionServer to Master in milliseconds.

Default: 3000


Sync the HLog to the HDFS after this interval if it has not accumulated enough entries to trigger a sync. Units: milliseconds.

Default: 1000


Limit for the number of regions after which no more region splitting should take place. This is not a hard limit for the number of regions but acts as a guideline for the regionserver to stop splitting after a certain limit. Default is MAX_INT; i.e. do not block splitting.

Default: 2147483647


Period at which we will roll the commit log regardless of how many edits it has.

Default: 3600000


The number of consecutive WAL close errors we will allow before triggering a server abort. A setting of 0 will cause the region server to abort if closing the current WAL writer fails during log rolling. Even a small value (2 or 3) will allow a region server to ride over transient HDFS errors.

Default: 2


The HLog file reader implementation.

Default: org.apache.hadoop.hbase.regionserver.wal.ProtobufLogReader


The HLog file writer implementation.

Default: org.apache.hadoop.hbase.regionserver.wal.ProtobufLogWriter

Maximum size of all memstores in a region server before new updates are blocked and flushes are forced. Defaults to 40% of heap. Updates are blocked and flushes are forced until size of all memstores in a region server hits

Default: 0.4

Maximum size of all memstores in a region server before flushes are forced. Defaults to 38% of heap. This value equal to causes the minimum possible flushing to occur when updates are blocked due to memstore limiting.

Default: 0.38


Maximum amount of time an edit lives in memory before being automatically flushed. Default 1 hour. Set it to 0 to disable automatic flushing.

Default: 3600000


Timeout value for the Catalog Janitor from the regionserver to META.

Default: 600000


The name of the Network Interface from which a region server should report its IP address.

Default: default


The host name or IP address of the name server (DNS) which a region server should use to determine the host name used by the master for communication and display purposes.

Default: default


A split policy determines when a region should be split. The various other split policies that are available currently are ConstantSizeRegionSplitPolicy, DisabledRegionSplitPolicy, DelimitedKeyPrefixRegionSplitPolicy, KeyPrefixRegionSplitPolicy etc.

Default: org.apache.hadoop.hbase.regionserver.IncreasingToUpperBoundRegionSplitPolicy


ZooKeeper session timeout in milliseconds. It is used in two different ways. First, this value is used in the ZK client that HBase uses to connect to the ensemble. It is also used by HBase when it starts a ZK server and it is passed as the 'maxSessionTimeout'. See For example, if a HBase region server connects to a ZK ensemble that's also managed by HBase, then the session timeout will be the one specified by this configuration. But, a region server that connects to an ensemble managed with a different configuration will be subjected that ensemble's maxSessionTimeout. So, even though HBase might propose using 90 seconds, the ensemble can have a max timeout lower than this and it will take precedence. The current default that ZK ships with is 40 seconds, which is lower than HBase's.

Default: 90000


Root ZNode for HBase in ZooKeeper. All of HBase's ZooKeeper files that are configured with a relative path will go under this node. By default, all of HBase's ZooKeeper file path are configured with a relative path, so they will all go under this directory unless changed.

Default: /hbase


Path to ZNode holding root region location. This is written by the master and read by clients and region servers. If a relative path is given, the parent folder will be ${zookeeper.znode.parent}. By default, this means the root location is stored at /hbase/root-region-server.

Default: root-region-server


Root ZNode for access control lists.

Default: acl


The name of the Network Interface from which a ZooKeeper server should report its IP address.

Default: default


The host name or IP address of the name server (DNS) which a ZooKeeper server should use to determine the host name used by the master for communication and display purposes.

Default: default


Port used by ZooKeeper peers to talk to each other. See for more information.

Default: 2888


Port used by ZooKeeper for leader election. See for more information.

Default: 3888


Instructs HBase to make use of ZooKeeper's multi-update functionality. This allows certain ZooKeeper operations to complete more quickly and prevents some issues with rare Replication failure scenarios (see the release note of HBASE-2611 for an example). IMPORTANT: only set this to true if all ZooKeeper servers in the cluster are on version 3.4+ and will not be downgraded. ZooKeeper versions before 3.4 do not support multi-update and will not fail gracefully if multi-update is invoked (see ZOOKEEPER-1495).

Default: false

Set to true to allow HBaseConfiguration to read the zoo.cfg file for ZooKeeper properties. Switching this to true is not recommended, since the functionality of reading ZK properties from a zoo.cfg file has been deprecated.

Default: false

Property from ZooKeeper's config zoo.cfg. The number of ticks that the initial synchronization phase can take.

Default: 10

Property from ZooKeeper's config zoo.cfg. The number of ticks that can pass between sending a request and getting an acknowledgment.

Default: 5

Property from ZooKeeper's config zoo.cfg. The directory where the snapshot is stored.

Default: ${hbase.tmp.dir}/zookeeper

Property from ZooKeeper's config zoo.cfg. The port at which the clients will connect.

Default: 2181

Property from ZooKeeper's config zoo.cfg. Limit on number of concurrent connections (at the socket level) that a single client, identified by IP address, may make to a single member of the ZooKeeper ensemble. Set high to avoid zk connection issues running standalone and pseudo-distributed.

Default: 300


Default size of the HTable client write buffer in bytes. A bigger buffer takes more memory -- on both the client and server side since server instantiates the passed write buffer to process it -- but a larger buffer size reduces the number of RPCs made. For an estimate of server-side memory-used, evaluate hbase.client.write.buffer * hbase.regionserver.handler.count

Default: 2097152


General client pause value. Used mostly as value to wait before running a retry of a failed get, region lookup, etc. See hbase.client.retries.number for description of how we backoff from this initial pause amount and how this pause works w/ retries.

Default: 100


Maximum retries. Used as maximum for all retryable operations such as the getting of a cell's value, starting a row update, etc. Retry interval is a rough function based on hbase.client.pause. At first we retry at this interval but then with backoff, we pretty quickly reach retrying every ten seconds. See HConstants#RETRY_BACKOFF for how the backup ramps up. Change this setting and hbase.client.pause to suit your workload.

Default: 35

The maximum number of concurrent tasks a single HTable instance will send to the cluster.

Default: 100


The maximum number of concurrent tasks a single HTable instance will send to a single region server.

Default: 5


The maximum number of concurrent connections the client will maintain to a single Region. That is, if there is already hbase.client.max.perregion.tasks writes in progress for this region, new puts won't be sent to this region until some writes finishes.

Default: 1


Number of rows that will be fetched when calling next on a scanner if it is not served from (local, client) memory. Higher caching values will enable faster scanners but will eat up more memory and some calls of next may take longer and longer times when the cache is empty. Do not set this value such that the time between invocations is greater than the scanner timeout; i.e. hbase.client.scanner.timeout.period

Default: 100


Specifies the combined maximum allowed size of a KeyValue instance. This is to set an upper boundary for a single entry saved in a storage file. Since they cannot be split it helps avoiding that a region cannot be split any further because the data is too large. It seems wise to set this to a fraction of the maximum region size. Setting it to zero or less disables the check.

Default: 10485760


Client scanner lease period in milliseconds.

Default: 60000


Default: 2


Maximum retries. This is maximum number of iterations to atomic bulk loads are attempted in the face of splitting operations 0 means never give up.

Default: 0


Period at which the region balancer runs in the Master.

Default: 300000


Rebalance if any regionserver has average + (average * slop) regions.

Default: 0.2


Time to sleep in between searches for work (in milliseconds). Used as sleep interval by service threads such as log roller.

Default: 10000


How many time to retry attempting to write a version file before just aborting. Each attempt is seperated by the hbase.server.thread.wakefrequency milliseconds.

Default: 3


Memstore will be flushed to disk if size of the memstore exceeds this number of bytes. Value is checked by a thread that runs every hbase.server.thread.wakefrequency.

Default: 134217728


If the memstores in a region are this size or larger when we go to close, run a "pre-flush" to clear out memstores before we put up the region closed flag and take the region offline. On close, a flush is run under the close flag to empty memory. During this time the region is offline and we are not taking on any writes. If the memstore content is large, this flush could take a long time to complete. The preflush is meant to clean out the bulk of the memstore before putting up the close flag and taking the region offline so the flush that runs under the close flag has little to do.

Default: 5242880


Block updates if memstore has hbase.hregion.block.memstore time hbase.hregion.flush.size bytes. Useful preventing runaway memstore during spikes in update traffic. Without an upper-bound, memstore fills such that when it flushes the resultant flush files take a long time to compact or split, or worse, we OOME.

Default: 2


Enables the MemStore-Local Allocation Buffer, a feature which works to prevent heap fragmentation under heavy write loads. This can reduce the frequency of stop-the-world GC pauses on large heaps.

Default: true


Maximum HStoreFile size. If any one of a column families' HStoreFiles has grown to exceed this value, the hosting HRegion is split in two.

Default: 10737418240


The time (in miliseconds) between 'major' compactions of all HStoreFiles in a region. Default: Set to 7 days. Major compactions tend to happen exactly when you need them least so enable them such that they run at off-peak for your deploy; or, since this setting is on a periodicity that is unlikely to match your loading, run the compactions via an external invocation out of a cron job or some such.

Default: 604800000


Jitter outer bound for major compactions. On each regionserver, we multiply the hbase.region.majorcompaction interval by some random fraction that is inside the bounds of this maximum. We then add this + or - product to when the next major compaction is to run. The idea is that major compaction does happen on every regionserver at exactly the same time. The smaller this number, the closer the compactions come together.

Default: 0.50


If more than this number of HStoreFiles in any one HStore (one HStoreFile is written per flush of memstore) then a compaction is run to rewrite all HStoreFiles files as one. Larger numbers put off compaction but when it runs, it takes longer to complete.

Default: 3


If more than this number of StoreFiles in any one Store (one StoreFile is written per flush of MemStore) then updates are blocked for this HRegion until a compaction is completed, or until hbase.hstore.blockingWaitTime has been exceeded.

Default: 10


The time an HRegion will block updates for after hitting the StoreFile limit defined by hbase.hstore.blockingStoreFiles. After this time has elapsed, the HRegion will stop blocking updates even if a compaction has not been completed.

Default: 90000


Max number of HStoreFiles to compact per 'minor' compaction.

Default: 10


How many KeyValues to read and then write in a batch when flushing or compacting. Do less if big KeyValues and problems with OOME. Do more if wide, small rows.

Default: 10

Enables StoreFileScanner parallel-seeking in StoreScanner, a feature which can reduce response latency under special conditions.

Default: false

The default thread pool size if parallel-seeking feature enabled.

Default: 10


Percentage of maximum heap (-Xmx setting) to allocate to block cache used by HFile/StoreFile. Default of 0.4 means allocate 40%. Set to 0 to disable but it's not recommended; you need at least enough cache to hold the storefile indices.

Default: 0.4


This allows to put non-root multi-level index blocks into the block cache at the time the index is being written.

Default: false


When the size of a leaf-level, intermediate-level, or root-level index block in a multi-level block index grows to this size, the block is written out and a new block is started.

Default: 131072


The HFile format version to use for new files. Set this to 1 to test backwards-compatibility. The default value of this option should be consistent with FixedFileTrailer.MAX_VERSION.

Default: 2


Enables cache-on-write for inline blocks of a compound Bloom filter.

Default: false


The size in bytes of a single block ("chunk") of a compound Bloom filter. This size is approximate, because Bloom blocks can only be inserted at data block boundaries, and the number of keys per data block varies.

Default: 131072

Whether an HFile block should be added to the block cache when the block is finished.

Default: false


Implementation of org.apache.hadoop.hbase.ipc.RpcServerEngine to be used for server RPC call marshalling.

Default: org.apache.hadoop.hbase.ipc.ProtobufRpcServerEngine


This is for the RPC layer to define how long HBase client applications take for a remote call to time out. It uses pings to check connections but will eventually throw a TimeoutException.

Default: 60000


This is another version of "hbase.rpc.timeout". For those RPC operation within cluster, we rely on this configuration to set a short timeout limitation for short operation. For example, short rpc timeout for region server's trying to report to active master can benefit quicker master failover process.

Default: 10000


Set no delay on rpc socket connections. See

Default: true


Full path to the kerberos keytab file to use for logging in the configured HMaster server principal.



Ex. "hbase/_HOST@EXAMPLE.COM". The kerberos principal name that should be used to run the HMaster process. The principal name should be in the form: user/hostname@DOMAIN. If "_HOST" is used as the hostname portion, it will be replaced with the actual hostname of the running instance.



Full path to the kerberos keytab file to use for logging in the configured HRegionServer server principal.



Ex. "hbase/_HOST@EXAMPLE.COM". The kerberos principal name that should be used to run the HRegionServer process. The principal name should be in the form: user/hostname@DOMAIN. If "_HOST" is used as the hostname portion, it will be replaced with the actual hostname of the running instance. An entry for this principal must exist in the file specified in hbase.regionserver.keytab.file



The policy configuration file used by RPC servers to make authorization decisions on client requests. Only used when HBase security is enabled.

Default: hbase-policy.xml


List of users or groups (comma-separated), who are allowed full privileges, regardless of stored ACLs, across the cluster. Only used when HBase security is enabled.



The update interval for master key for authentication tokens in servers in milliseconds. Only used when HBase security is enabled.

Default: 86400000


The maximum lifetime in milliseconds after which an authentication token expires. Only used when HBase security is enabled.

Default: 604800000


When a client is configured to attempt a secure connection, but attempts to connect to an insecure server, that server may instruct the client to switch to SASL SIMPLE (unsecure) authentication. This setting controls whether or not the client will accept this instruction from the server. When false (the default), the client will not allow the fallback to SIMPLE authentication, and will abort the connection.

Default: false


A comma-separated list of Coprocessors that are loaded by default on all tables. For any override coprocessor method, these classes will be called in order. After implementing your own Coprocessor, just put it in HBase's classpath and add the fully qualified class name here. A coprocessor can also be loaded on demand by setting HTableDescriptor.


The port for the HBase REST server.

Default: 8080

Defines the mode the REST server will be started in. Possible values are: false: All HTTP methods are permitted - GET/PUT/POST/DELETE. true: Only the GET method is permitted.

Default: false

The maximum number of threads of the REST server thread pool. Threads in the pool are reused to process REST requests. This controls the maximum number of requests processed concurrently. It may help to control the memory used by the REST server to avoid OOM issues. If the thread pool is full, incoming requests will be queued up and wait for some free threads.

Default: 100

The minimum number of threads of the REST server thread pool. The thread pool always has at least these number of threads so the REST server is ready to serve incoming requests.

Default: 2


Set to true to skip the 'hbase.defaults.for.version' check. Setting this to true can be useful in contexts other than the other side of a maven generation; i.e. running in an ide. You'll want to set this boolean to true to avoid seeing the RuntimException complaint: "hbase-default.xml file seems to be for and old version of HBase (\${hbase.version}), this version is X.X.X-SNAPSHOT"

Default: false


A comma-separated list of org.apache.hadoop.hbase.coprocessor.MasterObserver coprocessors that are loaded by default on the active HMaster process. For any implemented coprocessor methods, the listed classes will be called in order. After implementing your own MasterObserver, just put it in HBase's classpath and add the fully qualified class name here.



Set to true to cause the hosting server (master or regionserver) to abort if a coprocessor throws a Throwable object that is not IOException or a subclass of IOException. Setting it to true might be useful in development environments where one wants to terminate the server as soon as possible to simplify coprocessor failure analysis.

Default: false

Set true to enable online schema changes.

Default: true


Set to true to enable locking the table in zookeeper for schema change operations. Table locking from master prevents concurrent schema modifications to corrupt table state.

Default: true


The "core size" of the thread pool. New threads are created on every connection until this many threads are created.

Default: 16


The maximum size of the thread pool. When the pending request queue overflows, new threads are created until their number reaches this number. After that, the server starts dropping connections.

Default: 1000


The maximum number of pending Thrift connections waiting in the queue. If there are no idle threads in the pool, the server queues requests. Only when the queue overflows, new threads are added, up to hbase.thrift.maxQueuedRequests threads.

Default: 1000


The upper bound for the table pool used in the Thrift gateways server. Since this is per table name, we assume a single table and so with 1000 default worker threads max this is set to a matching number. For other workloads this number can be adjusted as needed.

Default: 1000


Use Thrift TFramedTransport on the server side. This is the recommended transport for thrift servers and requires a similar setting on the client side. Changing this to false will select the default transport, vulnerable to DoS when malformed requests are issued due to THRIFT-601.

Default: false


Default frame size when using framed transport

Default: 2


Use Thrift TCompactProtocol binary serialization protocol.

Default: false


The amount of off heap space to be allocated towards the experimental off heap cache. If you desire the cache to be disabled, simply set this value to 0.

Default: 0

Enable, if true, that file permissions should be assigned to the files written by the regionserver

Default: false

File permissions that should be used to write data files when is true

Default: 000


Whether to include the prefix "tbl.tablename" in per-column family metrics. If true, for each metric M, per-cf metrics will be reported for, if false, per-cf metrics will be aggregated by column-family across tables, and reported for cf.CF.M. In both cases, the aggregated metric M across tables and cfs will be reported.

Default: true


Whether to report metrics about time taken performing an operation on the region server. Get, Put, Delete, Increment, and Append can all have their times exposed through Hadoop metrics per CF and per region.

Default: true


Set to true to allow snapshots to be taken / restored / cloned.

Default: true


Set to true to take a snapshot before the restore operation. The snapshot taken will be used in case of failure, to restore the previous state. At the end of the restore operation this snapshot will be deleted

Default: true

Name of the failsafe snapshot taken by the restore operation. You can use the {}, {} and {restore.timestamp} variables to create a name based on what you are restoring.

Default: hbase-failsafe-{}-{restore.timestamp}


The number that determines how often we scan to see if compaction is necessary. Normally, compactions are done after some events (such as memstore flush), but if region didn't receive a lot of writes for some time, or due to different compaction policies, it may be necessary to check it periodically. The interval between checks is hbase.server.compactchecker.interval.multiplier multiplied by hbase.server.thread.wakefrequency.

Default: 1000

How long we wait on dfs lease recovery in total before giving up.

Default: 900000

How long between dfs recover lease invocations. Should be larger than the sum of the time it takes for the namenode to issue a block recovery command as part of datanode; dfs.heartbeat.interval and the time it takes for the primary datanode, performing block recovery to timeout on a dead datanode; usually dfs.socket.timeout. See the end of HBASE-8389 for more.

Default: 64000

If the DFSClient configuration is unset, we will use what is configured here as the short circuit read default direct byte buffer size. DFSClient native default is 1MB; HBase keeps its HDFS files open so number of file blocks * 1MB soon starts to add up and threaten OOME because of a shortage of direct memory. So, we set it down from the default. Make it > the default hbase block size set in the HColumnDescriptor which is usually 64k.

Default: 131072


If set to true, HBase will read data and then verify checksums for hfile blocks. Checksum verification inside HDFS will be switched off. If the hbase-checksum verification fails, then it will switch back to using HDFS checksums.

Default: true


Number of bytes in a newly created checksum chunk for HBase-level checksums in hfile blocks.

Default: 16384


Name of an algorithm that is used to compute checksums. Possible values are NULL, CRC32, CRC32C.

Default: CRC32


This setting activates the publication by the master of the status of the region server. When a region server dies and its recovery starts, the master will push this information to the client application, to let them cut the connection immediately instead of waiting for a timeout.

Default: false


Implementation of the status publication with a multicast message.

Default: org.apache.hadoop.hbase.master.ClusterStatusPublisher$MulticastPublisher


Implementation of the status listener with a multicast message.

Default: org.apache.hadoop.hbase.client.ClusterStatusListener$MulticastListener


Multicast address to use for the status publication by multicast.



Multicast port to use for the status publication by multicast.

Default: 6100


The directory from which the custom filter/co-processor jars can be loaded dynamically by the region server without the need to restart. However, an already loaded filter/co-processor class would not be un-loaded. See HBASE-1936 for more details.

Default: ${hbase.rootdir}/lib

Controls whether or not secure authentication is enabled for HBase. Possible values are 'simple' (no authentication), and 'kerberos'.

Default: simple

Servlet filters for REST service.



Class used to execute the regions balancing when the period occurs. See the class comment for more on how it works It replaces the DefaultLoadBalancer as the default (since renamed as the SimpleLoadBalancer).

Default: org.apache.hadoop.hbase.master.balancer.StochasticLoadBalancer


Set HBase environment variables in this file. Examples include options to pass the JVM on start of an HBase daemon such as heap size and garbarge collector configs. You can also set configurations for HBase configuration, log directories, niceness, ssh options, where to locate process pid files, etc. Open the file at conf/ and peruse its content. Each option is fairly well documented. Add your own environment variables here if you want them read by HBase daemons on startup.

Changes here will require a cluster restart for HBase to notice the change.


Edit this file to change rate at which HBase files are rolled and to change the level at which HBase logs messages.

Changes here will require a cluster restart for HBase to notice the change though log levels can be changed for particular daemons via the HBase UI.

2.3.4. Client configuration and dependencies connecting to an HBase cluster

If you are running HBase in standalone mode, you don't need to configure anything for your client to work provided that they are all on the same machine.

Since the HBase Master may move around, clients bootstrap by looking to ZooKeeper for current critical locations. ZooKeeper is where all these values are kept. Thus clients require the location of the ZooKeeper ensemble information before they can do anything else. Usually this the ensemble location is kept out in the hbase-site.xml and is picked up by the client from the CLASSPATH.

If you are configuring an IDE to run a HBase client, you should include the conf/ directory on your classpath so hbase-site.xml settings can be found (or add src/test/resources to pick up the hbase-site.xml used by tests).

Minimally, a client of HBase needs several libraries in its CLASSPATH when connecting to a cluster, including:

commons-configuration (commons-configuration-1.6.jar)
commons-lang (commons-lang-2.5.jar)
commons-logging (commons-logging-1.1.1.jar)
hadoop-core (hadoop-core-1.0.0.jar)
hbase (hbase-0.92.0.jar)
log4j (log4j-1.2.16.jar)
slf4j-api (slf4j-api-1.5.8.jar)
slf4j-log4j (slf4j-log4j12-1.5.8.jar)
zookeeper (zookeeper-3.4.2.jar)

An example basic hbase-site.xml for client only might look as follows:

<?xml version="1.0"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>
    <description>The directory shared by region servers.
</configuration> Java client configuration

The configuration used by a Java client is kept in an HBaseConfiguration instance. The factory method on HBaseConfiguration, HBaseConfiguration.create();, on invocation, will read in the content of the first hbase-site.xml found on the client's CLASSPATH, if one is present (Invocation will also factor in any hbase-default.xml found; an hbase-default.xml ships inside the hbase.X.X.X.jar). It is also possible to specify configuration directly without having to read from a hbase-site.xml. For example, to set the ZooKeeper ensemble for the cluster programmatically do as follows:

Configuration config = HBaseConfiguration.create();
config.set("hbase.zookeeper.quorum", "localhost");  // Here we are running zookeeper locally

If multiple ZooKeeper instances make up your ZooKeeper ensemble, they may be specified in a comma-separated list (just as in the hbase-site.xml file). This populated Configuration instance can then be passed to an HTable, and so on.

2.4. Example Configurations

2.4.1. Basic Distributed HBase Install

Here is an example basic configuration for a distributed ten node cluster. The nodes are named example0, example1, etc., through node example9 in this example. The HBase Master and the HDFS namenode are running on the node example0. RegionServers run on nodes example1-example9. A 3-node ZooKeeper ensemble runs on example1, example2, and example3 on the default ports. ZooKeeper data is persisted to the directory /export/zookeeper. Below we show what the main configuration files -- hbase-site.xml, regionservers, and -- found in the HBase conf directory might look like. hbase-site.xml

<?xml version="1.0"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>
    <description>The directory shared by RegionServers.
    <description>Property from ZooKeeper's config zoo.cfg.
    The directory where the snapshot is stored.
    <description>The directory shared by RegionServers.
    <description>The mode the cluster will be in. Possible values are
      false: standalone and pseudo-distributed setups with managed Zookeeper
      true: fully-distributed with unmanaged Zookeeper Quorum (see
</configuration> regionservers

In this file you list the nodes that will run RegionServers. In our case, these nodes are example1-example9.


Below we use a diff to show the differences from default in the file. Here we are setting the HBase heap to be 4G instead of the default 1G.

$ git diff
diff --git a/conf/ b/conf/
index e70ebc6..96f8c27 100644
--- a/conf/
+++ b/conf/
@@ -31,7 +31,7 @@ export JAVA_HOME=/usr/lib//jvm/java-6-sun/

 # The maximum amount of heap to use, in MB. Default is 1000.
-# export HBASE_HEAPSIZE=1000
+export HBASE_HEAPSIZE=4096

 # Extra Java runtime options.
 # Below are what we set by default.  May only work with SUN JVM.


Use rsync to copy the content of the conf directory to all nodes of the cluster.

2.5. The Important Configurations

Below we list what the important Configurations. We've divided this section into required configuration and worth-a-look recommended configs.

2.5.1. Required Configurations

Review the Section 2.1.2, “Operating System” and Section 2.1.3, “Hadoop” sections. Big Cluster Configurations

If a cluster with a lot of regions, it is possible if an eager beaver regionserver checks in soon after master start while all the rest in the cluster are laggardly, this first server to checkin will be assigned all regions. If lots of regions, this first server could buckle under the load. To prevent the above scenario happening up the hbase.master.wait.on.regionservers.mintostart from its default value of 1. See HBASE-6389 Modify the conditions to ensure that Master waits for sufficient number of Region Servers before starting region assignments for more detail. If a backup Master, making primary Master fail fast

If the primary Master loses its connection with ZooKeeper, it will fall into a loop where it keeps trying to reconnect. Disable this functionality if you are running more than one Master: i.e. a backup Master. Failing to do so, the dying Master may continue to receive RPCs though another Master has assumed the role of primary. See the configuration

2.5.2. Recommended Configurations ZooKeeper Configuration zookeeper.session.timeout

The default timeout is three minutes (specified in milliseconds). This means that if a server crashes, it will be three minutes before the Master notices the crash and starts recovery. You might like to tune the timeout down to a minute or even less so the Master notices failures the sooner. Before changing this value, be sure you have your JVM garbage collection configuration under control otherwise, a long garbage collection that lasts beyond the ZooKeeper session timeout will take out your RegionServer (You might be fine with this -- you probably want recovery to start on the server if a RegionServer has been in GC for a long period of time).

To change this configuration, edit hbase-site.xml, copy the changed file around the cluster and restart.

We set this value high to save our having to field noob questions up on the mailing lists asking why a RegionServer went down during a massive import. The usual cause is that their JVM is untuned and they are running into long GC pauses. Our thinking is that while users are getting familiar with HBase, we'd save them having to know all of its intricacies. Later when they've built some confidence, then they can play with configuration such as this. Number of ZooKeeper Instances

See Chapter 17, ZooKeeper. HDFS Configurations dfs.datanode.failed.volumes.tolerated

This is the "...number of volumes that are allowed to fail before a datanode stops offering service. By default any volume failure will cause a datanode to shutdown" from the hdfs-default.xml description. If you have > three or four disks, you might want to set this to 1 or if you have many disks, two or more. hbase.regionserver.handler.count

This setting defines the number of threads that are kept open to answer incoming requests to user tables. The rule of thumb is to keep this number low when the payload per request approaches the MB (big puts, scans using a large cache) and high when the payload is small (gets, small puts, ICVs, deletes). The total size of the queries in progress is limited by the setting "ipc.server.max.callqueue.size".

It is safe to set that number to the maximum number of incoming clients if their payload is small, the typical example being a cluster that serves a website since puts aren't typically buffered and most of the operations are gets.

The reason why it is dangerous to keep this setting high is that the aggregate size of all the puts that are currently happening in a region server may impose too much pressure on its memory, or even trigger an OutOfMemoryError. A region server running on low memory will trigger its JVM's garbage collector to run more frequently up to a point where GC pauses become noticeable (the reason being that all the memory used to keep all the requests' payloads cannot be trashed, no matter how hard the garbage collector tries). After some time, the overall cluster throughput is affected since every request that hits that region server will take longer, which exacerbates the problem even more.

You can get a sense of whether you have too little or too many handlers by Section, “Enabling RPC-level logging” on an individual RegionServer then tailing its logs (Queued requests consume memory). Configuration for large memory machines

HBase ships with a reasonable, conservative configuration that will work on nearly all machine types that people might want to test with. If you have larger machines -- HBase has 8G and larger heap -- you might the following configuration options helpful. TODO. Compression

You should consider enabling ColumnFamily compression. There are several options that are near-frictionless and in most all cases boost performance by reducing the size of StoreFiles and thus reducing I/O.

See Appendix C, Compression In HBase for more information. Configuring the size and number of WAL files

HBase uses Section 9.6.5, “Write Ahead Log (WAL)” to recover the memstore data that has not been flushed to disk in case of an RS failure. These WAL files should be configured to be slightly smaller than HDFS block (by default, HDFS block is 64Mb and WAL file is ~60Mb).

HBase also has a limit on number of WAL files, designed to ensure there's never too much data that needs to be replayed during recovery. This limit needs to be set according to memstore configuration, so that all the necessary data would fit. It is recommended to allocated enough WAL files to store at least that much data (when all memstores are close to full). For example, with 16Gb RS heap, default memstore settings (0.4), and default WAL file size (~60Mb), 16Gb*0.4/60, the starting point for WAL file count is ~109. However, as all memstores are not expected to be full all the time, less WAL files can be allocated. Managed Splitting

Rather than let HBase auto-split your Regions, manage the splitting manually [12]. With growing amounts of data, splits will continually be needed. Since you always know exactly what regions you have, long-term debugging and profiling is much easier with manual splits. It is hard to trace the logs to understand region level problems if it keeps splitting and getting renamed. Data offlining bugs + unknown number of split regions == oh crap! If an HLog or StoreFile was mistakenly unprocessed by HBase due to a weird bug and you notice it a day or so later, you can be assured that the regions specified in these files are the same as the current regions and you have less headaches trying to restore/replay your data. You can finely tune your compaction algorithm. With roughly uniform data growth, it's easy to cause split / compaction storms as the regions all roughly hit the same data size at the same time. With manual splits, you can let staggered, time-based major compactions spread out your network IO load.

How do I turn off automatic splitting? Automatic splitting is determined by the configuration value hbase.hregion.max.filesize. It is not recommended that you set this to Long.MAX_VALUE in case you forget about manual splits. A suggested setting is 100GB, which would result in > 1hr major compactions if reached.

What's the optimal number of pre-split regions to create? Mileage will vary depending upon your application. You could start low with 10 pre-split regions / server and watch as data grows over time. It's better to err on the side of too little regions and rolling split later. A more complicated answer is that this depends upon the largest storefile in your region. With a growing data size, this will get larger over time. You want the largest region to be just big enough that the Store compact selection algorithm only compacts it due to a timed major. If you don't, your cluster can be prone to compaction storms as the algorithm decides to run major compactions on a large series of regions all at once. Note that compaction storms are due to the uniform data growth, not the manual split decision.

If you pre-split your regions too thin, you can increase the major compaction interval by configuring HConstants.MAJOR_COMPACTION_PERIOD. If your data size grows too large, use the (post-0.90.0 HBase) org.apache.hadoop.hbase.util.RegionSplitter script to perform a network IO safe rolling split of all regions. Managed Compactions

A common administrative technique is to manage major compactions manually, rather than letting HBase do it. By default, HConstants.MAJOR_COMPACTION_PERIOD is one day and major compactions may kick in when you least desire it - especially on a busy system. To turn off automatic major compactions set the value to 0.

It is important to stress that major compactions are absolutely necessary for StoreFile cleanup, the only variant is when they occur. They can be administered through the HBase shell, or via HBaseAdmin.

For more information about compactions and the compaction file selection process, see Section, “Compaction” Speculative Execution

Speculative Execution of MapReduce tasks is on by default, and for HBase clusters it is generally advised to turn off Speculative Execution at a system-level unless you need it for a specific case, where it can be configured per-job. Set the properties and mapred.reduce.tasks.speculative.execution to false.

2.5.3. Other Configurations Balancer

The balancer is a periodic operation which is run on the master to redistribute regions on the cluster. It is configured via hbase.balancer.period and defaults to 300000 (5 minutes).

See Section, “LoadBalancer” for more information on the LoadBalancer. Disabling Blockcache

Do not turn off block cache (You'd do it by setting hbase.block.cache.size to zero). Currently we do not do well if you do this because the regionserver will spend all its time loading hfile indices over and over again. If your working set it such that block cache does you no good, at least size the block cache such that hfile indices will stay up in the cache (you can get a rough idea on the size you need by surveying regionserver UIs; you'll see index block size accounted near the top of the webpage). Nagle's or the small package problem

If a big 40ms or so occasional delay is seen in operations against HBase, try the Nagles' setting. For example, see the user mailing list thread, Inconsistent scan performance with caching set to 1 and the issue cited therein where setting notcpdelay improved scan speeds. You might also see the graphs on the tail of HBASE-7008 Set scanner caching to a better default where our Lars Hofhansl tries various data sizes w/ Nagle's on and off measuring the effect. Better Mean Time to Recover (MTTR)

See the Deveraj Das an Nicolas Liochon blog post Introduction to HBase Mean Time to Recover (MTTR) for a brief introduction. The issue HBASE-8354 forces Namenode into loop with lease recovery requests is messy but has a bunch of good discussion toward the end on low timeouts and how to effect faster recovery including citation of fixes added to HDFS. Read the Varun Sharma comments.

[2] Be careful editing XML. Make sure you close all elements. Run your file through xmllint or similar to ensure well-formedness of your document after an edit session.

[3] The hadoop-dns-checker tool can be used to verify DNS is working correctly on the cluster. The project README file provides detailed instructions on usage.

[4] See Jack Levin's major hdfs issues note up on the user list.

[5] The requirement that a database requires upping of system limits is not peculiar to Apache HBase. See for example the section Setting Shell Limits for the Oracle User in Short Guide to install Oracle 10 on Linux.

[6] A useful read setting config on you hadoop cluster is Aaron Kimballs' Configuration Parameters: What can you just ignore?

[8] The Cloudera blog post An update on Apache Hadoop 1.0 by Charles Zedlweski has a nice exposition on how all the Hadoop versions relate. Its worth checking out if you are having trouble making sense of the Hadoop version morass.

[9] See Hadoop HDFS: Deceived by Xciever for an informative rant on xceivering.

[10] The pseudo-distributed vs fully-distributed nomenclature comes from Hadoop.

[11] See Section, “Pseudo-distributed Extras” for notes on how to start extra Masters and RegionServers when running pseudo-distributed.

[12] What follows is taken from the javadoc at the head of the org.apache.hadoop.hbase.util.RegionSplitter tool added to HBase post-0.90.0 release.

Chapter 3. Upgrading

You cannot skip major verisons upgrading. If you are upgrading from version 0.90.x to 0.94.x, you must first go from 0.90.x to 0.92.x and then go from 0.92.x to 0.94.x.

Review Chapter 2, Apache HBase Configuration, in particular the section on Hadoop version.

3.1. HBase version numbers

HBase has not walked a straight line where version numbers are concerned. Since we came up out of hadoop itself, we originally tracked hadoop versioning. Later we left hadoop versioning behind because we were moving at a different rate to that of our parent. If you are into the arcane, checkout our old wiki page on HBase Versioning which tries to connect the HBase version dots.

3.1.1. Odd/Even Versioning or "Development"" Series Releases

Ahead of big releases, we have been putting up preview versions to start the feedback cycle turning-over earlier. These "Development" Series releases, always odd-numbered, come with no guarantees, not even regards being able to upgrade between two sequential releases (we reserve the right to break compatibility across "Development" Series releases). Needless to say, these releases are not for production deploys. They are a preview of what is coming in the hope that interested parties will take the release for a test drive and flag us early if we there are issues we've missed ahead of our rolling a production-worthy release.

Our first "Development" Series was the 0.89 set that came out ahead of HBase 0.90.0. HBase 0.95 is another "Development" Series that portends HBase 0.96.0.

3.1.2. Binary Compatibility

When we say two HBase versions are compatible, we mean that the versions are wire and binary compatible. Compatible HBase versions means that clients can talk to compatible but differently versioned servers. It means too that you can just swap out the jars of one version and replace them with the jars of another, compatible version and all will just work. Unless otherwise specified, HBase point versions are binary compatible. You can safely do rolling upgrades between binary compatible versions; i.e. across point versions: e.g. from 0.94.5 to 0.94.6[13].

3.1.3. Rolling Upgrade between versions/Binary compatibility

Unless otherwise specified, HBase point versions are binary compatible. you can do a rolling upgrade between hbase point versions; for example, you can go to 0.94.6 from 0.94.5 by doing a rolling upgrade across the cluster replacing the 0.94.5 binary with a 0.94.6 binary.

3.2. Upgrading from 0.94.x to 0.96.x

The Singularity

You will have to stop your old 0.94.x cluster completely to upgrade. If you are replicating between clusters, both clusters will have to go down to upgrade. Make sure it is a clean shutdown. The less WAL files around, the faster the upgrade will run (the upgrade will split any log files it finds in the filesystem as part of the upgrade process). All clients must be upgraded to 0.96 too.

The API has changed. You will need to recompile your code against 0.96 and you may need to adjust applications to go against new APIs (TODO: List of changes).

3.2.1. Executing the 0.96 Upgrade


HDFS and ZooKeeper should be up and running during the upgrade process.

hbase-0.96.0 comes with an upgrade script. Run

$ bin/hbase upgrade

to see its usage. The script has two main modes: -check, and -execute. check

The check step is run against a running 0.94 cluster. Run it from a downloaded 0.96.x binary. The check step is looking for the presence of HFileV1 files. These are unsupported in hbase-0.96.0. To purge them -- have them rewritten as HFileV2 -- you must run a compaction.

The check step prints stats at the end of its run (grep for “Result:” in the log) printing absolute path of the tables it scanned, any HFileV1 files found, the regions containing said files (the regions we need to major compact to purge the HFileV1s), and any corrupted files if any found. A corrupt file is unreadable, and so is undefined (neither HFileV1 nor HFileV2).

To run the check step, run

$ bin/hbase upgrade -check

. Here is sample output:

             Tables Processed:

             Count of HFileV1: 2
             hdfs://localhost:41020/myHBase/usertable    /fa02dac1f38d03577bd0f7e666f12812/family/249450144068442524
             hdfs://localhost:41020/myHBase/usertable    /ecdd3eaee2d2fcf8184ac025555bb2af/family/249450144068442512

             Count of corrupted files: 1
             Corrupted Files:
             Count of Regions with HFileV1: 2
             Regions to Major Compact:

             There are some HFileV1, or corrupt files (files with incorrect major version)

In the above sample output, there are two HFileV1 in two regions, and one corrupt file. Corrupt files should probably be removed. The regions that have HFileV1s need to be major compacted. To major compact, start up the hbase shell and review how to compact an individual region. After the major compaction is done, rerun the check step and the HFileV1s shoudl be gone, replaced by HFileV2 instances.

By default, the check step scans the hbase root directory (defined as hbase.rootdir in the configuration). To scan a specific directory only, pass the -dir option.

$ bin/hbase upgrade -check -dir /myHBase/testTable

The above command would detect HFileV1s in the /myHBase/testTable directory.

Once the check step reports all the HFileV1 files have been rewritten, it is safe to proceed with the upgrade. execute

After the check step shows the cluster is free of HFileV1, it is safe to proceed with the upgrade. Next is the execute step. You must SHUTDOWN YOUR 0.94.x CLUSTER before you can run the execute step. The execute step will not run if it detects running HBase masters or regionservers.


HDFS and ZooKeeper should be up and running during the upgrade process. If zookeeper is managed by HBase, then you can start zookeeper so it is available to the upgrade by running

$ ./hbase/bin/  start zookeeper

The execute upgrade step is made of three substeps.

  • Namespaces: HBase 0.96.0 has support for namespaces. The upgrade needs to reorder directories in the filesystem for namespaces to work.

  • ZNodes: All znodes are purged so that new ones can be written in their place using a new protobuf'ed format and a few are migrated in place: e.g. replication and table state znodes

  • WAL Log Splitting: If the 0.94.x cluster shutdown was not clean, we'll split WAL logs as part of migration before we startup on 0.96.0. This WAL splitting runs slower than the native distributed WAL splitting because it is all inside the single upgrade process (so try and get a clean shutdown of the 0.94.0 cluster if you can).

To run the execute step, make sure that first you have copied hbase-0.96.0 binaries everywhere under servers and under clients. Make sure the 0.94.0 cluster is down. Then do as follows:

$ bin/hbase upgrade -execute

Here is some sample output

             Starting Namespace upgrade
             Created version file at hdfs://localhost:41020/myHBase with version=7
             Migrating table testTable to hdfs://localhost:41020/myHBase/.data/default/testTable
             Created version file at hdfs://localhost:41020/myHBase with version=8
             Successfully completed NameSpace upgrade.
             Starting Znode upgrade
             Successfully completed Znode upgrade

             Starting Log splitting
             Successfully completed Log splitting

If the output from the execute step looks good, start hbase-0.96.0. Troubleshooting Old Client connecting to 0.96 cluster

It will fail with an exception like the below. Upgrade.

17:22:15  Exception in thread "main" java.lang.IllegalArgumentException: Not a host:port pair: PBUF
17:22:15  *
17:22:15 ��  ���(
17:22:15    at org.apache.hadoop.hbase.util.Addressing.parseHostname(
17:22:15    at org.apache.hadoop.hbase.ServerName.&init>(
17:22:15    at org.apache.hadoop.hbase.ServerName.parseVersionedServerName(
17:22:15    at org.apache.hadoop.hbase.MasterAddressTracker.bytesToServerName(
17:22:15    at org.apache.hadoop.hbase.MasterAddressTracker.getMasterAddress(
17:22:15    at org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation.getMaster(
17:22:15    at org.apache.hadoop.hbase.client.HBaseAdmin.&init>(
17:22:15    at Client_4_3_0.setup(
17:22:15    at Client_4_3_0.main(

3.3. Upgrading from 0.92.x to 0.94.x

We used to think that 0.92 and 0.94 were interface compatible and that you can do a rolling upgrade between these versions but then we figured that HBASE-5357 Use builder pattern in HColumnDescriptor changed method signatures so rather than return void they instead return HColumnDescriptor. This will throw

java.lang.NoSuchMethodError: org.apache.hadoop.hbase.HColumnDescriptor.setMaxVersions(I)V

.... so 0.92 and 0.94 are NOT compatible. You cannot do a rolling upgrade between them.

3.4. Upgrading from 0.90.x to 0.92.x

Upgrade Guide

You will find that 0.92.0 runs a little differently to 0.90.x releases. Here are a few things to watch out for upgrading from 0.90.x to 0.92.0.


If you've not patience, here are the important things to know upgrading.

  1. Once you upgrade, you can’t go back.
  2. MSLAB is on by default. Watch that heap usage if you have a lot of regions.
  3. Distributed splitting is on by defaul. It should make region server failover faster.
  4. There’s a separate tarball for security.
  5. If -XX:MaxDirectMemorySize is set in your, it’s going to enable the experimental off-heap cache (You may not want this).

3.4.1. You can’t go back!

To move to 0.92.0, all you need to do is shutdown your cluster, replace your hbase 0.90.x with hbase 0.92.0 binaries (be sure you clear out all 0.90.x instances) and restart (You cannot do a rolling restart from 0.90.x to 0.92.x -- you must restart). On startup, the .META. table content is rewritten removing the table schema from the info:regioninfo column. Also, any flushes done post first startup will write out data in the new 0.92.0 file format, HFile V2. This means you cannot go back to 0.90.x once you’ve started HBase 0.92.0 over your HBase data directory.

3.4.2. MSLAB is ON by default

In 0.92.0, the hbase.hregion.memstore.mslab.enabled flag is set to true (See Section, “Long GC pauses”). In 0.90.x it was false. When it is enabled, memstores will step allocate memory in MSLAB 2MB chunks even if the memstore has zero or just a few small elements. This is fine usually but if you had lots of regions per regionserver in a 0.90.x cluster (and MSLAB was off), you may find yourself OOME'ing on upgrade because the thousands of regions * number of column families * 2MB MSLAB (at a minimum) puts your heap over the top. Set hbase.hregion.memstore.mslab.enabled to false or set the MSLAB size down from 2MB by setting hbase.hregion.memstore.mslab.chunksize to something less.

3.4.3. Distributed splitting is on by default

Previous, WAL logs on crash were split by the Master alone. In 0.92.0, log splitting is done by the cluster (See See “HBASE-1364 [performance] Distributed splitting of regionserver commit logs”). This should cut down significantly on the amount of time it takes splitting logs and getting regions back online again.

3.4.4. Memory accounting is different now

In 0.92.0, Appendix E, HFile format version 2 indices and bloom filters take up residence in the same LRU used caching blocks that come from the filesystem. In 0.90.x, the HFile v1 indices lived outside of the LRU so they took up space even if the index was on a ‘cold’ file, one that wasn’t being actively used. With the indices now in the LRU, you may find you have less space for block caching. Adjust your block cache accordingly. See the Section 9.6.4, “Block Cache” for more detail. The block size default size has been changed in 0.92.0 from 0.2 (20 percent of heap) to 0.25.

3.4.5. On the Hadoop version to use

Run 0.92.0 on Hadoop 1.0.x (or CDH3u3 when it ships). The performance benefits are worth making the move. Otherwise, our Hadoop prescription is as it has been; you need an Hadoop that supports a working sync. See Section 2.1.3, “Hadoop”.

If running on Hadoop 1.0.x (or CDH3u3), enable local read. See Practical Caching presentation for ruminations on the performance benefits ‘going local’ (and for how to enable local reads).

3.4.6. HBase 0.92.0 ships with ZooKeeper 3.4.2

If you can, upgrade your zookeeper. If you can’t, 3.4.2 clients should work against 3.3.X ensembles (HBase makes use of 3.4.2 API).

3.4.7. Online alter is off by default

In 0.92.0, we’ve added an experimental online schema alter facility (See Its off by default. Enable it at your own risk. Online alter and splitting tables do not play well together so be sure your cluster quiescent using this feature (for now).

3.4.8. WebUI

The webui has had a few additions made in 0.92.0. It now shows a list of the regions currently transitioning, recent compactions/flushes, and a process list of running processes (usually empty if all is well and requests are being handled promptly). Other additions including requests by region, a debugging servlet dump, etc.

3.4.9. Security tarball

We now ship with two tarballs; secure and insecure HBase. Documentation on how to setup a secure HBase is on the way.

3.4.10. Experimental off-heap cache

A new cache was contributed to 0.92.0 to act as a solution between using the “on-heap” cache which is the current LRU cache the region servers have and the operating system cache which is out of our control. To enable, set “-XX:MaxDirectMemorySize” in to the value for maximum direct memory size and specify hbase.offheapcache.percentage in hbase-site.xml with the percentage that you want to dedicate to off-heap cache. This should only be set for servers and not for clients. Use at your own risk. See this blog post for additional information on this new experimental feature:

3.4.11. Changes in HBase replication

0.92.0 adds two new features: multi-slave and multi-master replication. The way to enable this is the same as adding a new peer, so in order to have multi-master you would just run add_peer for each cluster that acts as a master to the other slave clusters. Collisions are handled at the timestamp level which may or may not be what you want, this needs to be evaluated on a per use case basis. Replication is still experimental in 0.92 and is disabled by default, run it at your own risk.

3.4.12. RegionServer now aborts if OOME

If an OOME, we now have the JVM kill -9 the regionserver process so it goes down fast. Previous, a RegionServer might stick around after incurring an OOME limping along in some wounded state. To disable this facility, and recommend you leave it in place, you’d need to edit the bin/hbase file. Look for the addition of the -XX:OnOutOfMemoryError="kill -9 %p" arguments (See [HBASE-4769] - ‘Abort RegionServer Immediately on OOME’)

3.4.13. HFile V2 and the “Bigger, Fewer” Tendency

0.92.0 stores data in a new format, Appendix E, HFile format version 2. As HBase runs, it will move all your data from HFile v1 to HFile v2 format. This auto-migration will run in the background as flushes and compactions run. HFile V2 allows HBase run with larger regions/files. In fact, we encourage that all HBasers going forward tend toward Facebook axiom #1, run with larger, fewer regions. If you have lots of regions now -- more than 100s per host -- you should look into setting your region size up after you move to 0.92.0 (In 0.92.0, default size is now 1G, up from 256M), and then running online merge tool (See “HBASE-1621 merge tool should work on online cluster, but disabled table”).

3.5. Upgrading to HBase 0.90.x from 0.20.x or 0.89.x

This version of 0.90.x HBase can be started on data written by HBase 0.20.x or HBase 0.89.x. There is no need of a migration step. HBase 0.89.x and 0.90.x does write out the name of region directories differently -- it names them with a md5 hash of the region name rather than a jenkins hash -- so this means that once started, there is no going back to HBase 0.20.x.

Be sure to remove the hbase-default.xml from your conf directory on upgrade. A 0.20.x version of this file will have sub-optimal configurations for 0.90.x HBase. The hbase-default.xml file is now bundled into the HBase jar and read from there. If you would like to review the content of this file, see it in the src tree at src/main/resources/hbase-default.xml or see Section, “HBase Default Configuration”.

Finally, if upgrading from 0.20.x, check your .META. schema in the shell. In the past we would recommend that users run with a 16kb MEMSTORE_FLUSHSIZE. Run hbase> scan '-ROOT-' in the shell. This will output the current .META. schema. Check MEMSTORE_FLUSHSIZE size. Is it 16kb (16384)? If so, you will need to change this (The 'normal'/default value is 64MB (67108864)). Run the script bin/set_meta_memstore_size.rb. This will make the necessary edit to your .META. schema. Failure to run this change will make for a slow cluster [14] .

Chapter 4. The Apache HBase Shell

The Apache HBase Shell is (J)Ruby's IRB with some HBase particular commands added. Anything you can do in IRB, you should be able to do in the HBase Shell.

To run the HBase shell, do as follows:

$ ./bin/hbase shell

Type help and then <RETURN> to see a listing of shell commands and options. Browse at least the paragraphs at the end of the help emission for the gist of how variables and command arguments are entered into the HBase shell; in particular note how table names, rows, and columns, etc., must be quoted.

See Section 1.2.3, “Shell Exercises” for example basic shell operation.

Here is a nicely formatted listing of all shell commands by Rajeshbabu Chintaguntla.

4.1. Scripting

For examples scripting Apache HBase, look in the HBase bin directory. Look at the files that end in *.rb. To run one of these files, do as follows:

$ ./bin/hbase org.jruby.Main PATH_TO_SCRIPT

4.2. Shell Tricks

4.2.1. Table variables

HBase 0.95 adds shell commands that provide a jruby-style object-oriented references for tables. Previously all of the shell commands that act upon a table have a procedural style that always took the name of the table as an argument. HBase 0.95 introduces the ability to assign a table to a jruby variable. The table reference can be used to perform data read write operations such as puts, scans, and gets well as admin functionality such as disabling, dropping, describing tables.

For example, previously you would always specify a table name:

hbase(main):000:0> create ‘t’, ‘f’
0 row(s) in 1.0970 seconds
hbase(main):001:0> put 't', 'rold', 'f', 'v'
0 row(s) in 0.0080 seconds

hbase(main):002:0> scan 't' 
ROW                                COLUMN+CELL                                                                                      
 rold                              column=f:, timestamp=1378473207660, value=v                                                      
1 row(s) in 0.0130 seconds

hbase(main):003:0> describe 't'
DESCRIPTION                                                                           ENABLED                                       
 't', {NAME => 'f', DATA_BLOCK_ENCODING => 'NONE', BLOOMFILTER => 'ROW', REPLICATION_ true                                          
 SCOPE => '0', VERSIONS => '1', COMPRESSION => 'NONE', MIN_VERSIONS => '0', TTL => '2                                               
 147483647', KEEP_DELETED_CELLS => 'false', BLOCKSIZE => '65536', IN_MEMORY => 'false                                               
 ', BLOCKCACHE => 'true'}                                                                                 
1 row(s) in 1.4430 seconds

hbase(main):004:0> disable 't'
0 row(s) in 14.8700 seconds

hbase(main):005:0> drop 't'
0 row(s) in 23.1670 seconds


Now you can assign the table to a variable and use the results in jruby shell code.

hbase(main):007 > t = create 't', 'f'
0 row(s) in 1.0970 seconds

=> Hbase::Table - t
hbase(main):008 > t.put 'r', 'f', 'v'
0 row(s) in 0.0640 seconds
hbase(main):009 > t.scan
ROW                           COLUMN+CELL                                                                        
 r                            column=f:, timestamp=1331865816290, value=v                                        
1 row(s) in 0.0110 seconds
hbase(main):010:0> t.describe
DESCRIPTION                                                                           ENABLED                                       
 't', {NAME => 'f', DATA_BLOCK_ENCODING => 'NONE', BLOOMFILTER => 'ROW', REPLICATION_ true                                          
 SCOPE => '0', VERSIONS => '1', COMPRESSION => 'NONE', MIN_VERSIONS => '0', TTL => '2                                               
 147483647', KEEP_DELETED_CELLS => 'false', BLOCKSIZE => '65536', IN_MEMORY => 'false                                               
 ', BLOCKCACHE => 'true'}                                                                                 
1 row(s) in 0.0210 seconds
hbase(main):038:0> t.disable
0 row(s) in 6.2350 seconds
hbase(main):039:0> t.drop
0 row(s) in 0.2340 seconds

If the table has already been created, you can assign a Table to a variable by using the get_table method:

hbase(main):011 > create 't','f'
0 row(s) in 1.2500 seconds

=> Hbase::Table - t
hbase(main):012:0> tab = get_table 't'
0 row(s) in 0.0010 seconds

=> Hbase::Table - t
hbase(main):013:0> tab.put ‘r1’ ,’f’, ‘v’ 
0 row(s) in 0.0100 seconds
hbase(main):014:0> tab.scan
ROW                                COLUMN+CELL                                                                                      
 r1                                column=f:, timestamp=1378473876949, value=v                                                      
1 row(s) in 0.0240 seconds

The list functionality has also been extended so that it returns a list of table names as strings. You can then use jruby to script table operations based on these names. The list_snapshots command also acts similarly.

hbase(main):016 > tables = list(‘t.*’)
1 row(s) in 0.1040 seconds

=> #<#<Class:0x7677ce29>:0x21d377a4>
hbase(main):017:0> { |t| disable t ; drop  t}
0 row(s) in 2.2510 seconds

=> [nil]

4.2.2. irbrc

Create an .irbrc file for yourself in your home directory. Add customizations. A useful one is command history so commands are save across Shell invocations:

                        $ more .irbrc
                        require 'irb/ext/save-history'
                        IRB.conf[:SAVE_HISTORY] = 100
                        IRB.conf[:HISTORY_FILE] = "#{ENV['HOME']}/.irb-save-history"

See the ruby documentation of .irbrc to learn about other possible confiurations.

4.2.3. LOG data to timestamp

To convert the date '08/08/16 20:56:29' from an hbase log into a timestamp, do:

                    hbase(main):021:0> import java.text.SimpleDateFormat
                    hbase(main):022:0> import java.text.ParsePosition
                    hbase(main):023:0>"yy/MM/dd HH:mm:ss").parse("08/08/16 20:56:29", => 1218920189000

To go the other direction:

                    hbase(main):021:0> import java.util.Date
                    hbase(main):022:0> => "Sat Aug 16 20:56:29 UTC 2008"

To output in a format that is exactly like that of the HBase log format will take a little messing with SimpleDateFormat.

4.2.4. Debug Shell debug switch

You can set a debug switch in the shell to see more output -- e.g. more of the stack trace on exception -- when you run a command:

hbase> debug <RETURN> DEBUG log level

To enable DEBUG level logging in the shell, launch it with the -d option.

$ ./bin/hbase shell -d

4.2.5. Commands count

Count command returns the number of rows in a table. It's quite fast when configured with the right CACHE

hbase> count '<tablename>', CACHE => 1000

The above count fetches 1000 rows at a time. Set CACHE lower if your rows are big. Default is to fetch one row at a time.

Chapter 5. Data Model

In short, applications store data into an HBase table. Tables are made of rows and columns. All columns in HBase belong to a particular column family. Table cells -- the intersection of row and column coordinates -- are versioned. A cell’s content is an uninterpreted array of bytes.

Table row keys are also byte arrays so almost anything can serve as a row key from strings to binary representations of longs or even serialized data structures. Rows in HBase tables are sorted by row key. The sort is byte-ordered. All table accesses are via the table row key -- its primary key.

5.1. Conceptual View

The following example is a slightly modified form of the one on page 2 of the BigTable paper. There is a table called webtable that contains two column families named contents and anchor. In this example, anchor contains two columns (, and contents contains one column (contents:html).

Column Names

By convention, a column name is made of its column family prefix and a qualifier. For example, the column contents:html is of the column family contents The colon character (:) delimits the column family from the column family qualifier.

Table 5.1. Table webtable

Row KeyTime StampColumnFamily contentsColumnFamily anchor
"com.cnn.www"t9 = "CNN"
"com.cnn.www"t8 = ""
"com.cnn.www"t6contents:html = "<html>..." 
"com.cnn.www"t5contents:html = "<html>..." 
"com.cnn.www"t3contents:html = "<html>..." 

5.2. Physical View

Although at a conceptual level tables may be viewed as a sparse set of rows. Physically they are stored on a per-column family basis. New columns (i.e., columnfamily:column) can be added to any column family without pre-announcing them.

Table 5.2. ColumnFamily anchor

Row KeyTime StampColumn Family anchor
"com.cnn.www" = "CNN"
"com.cnn.www" = ""

Table 5.3. ColumnFamily contents

Row KeyTime StampColumnFamily "contents:"
"com.cnn.www"t6contents:html = "<html>..."
"com.cnn.www"t5contents:html = "<html>..."
"com.cnn.www"t3contents:html = "<html>..."

It is important to note in the diagram above that the empty cells shown in the conceptual view are not stored since they need not be in a column-oriented storage format. Thus a request for the value of the contents:html column at time stamp t8 would return no value. Similarly, a request for an value at time stamp t9 would return no value. However, if no timestamp is supplied, the most recent value for a particular column would be returned and would also be the first one found since timestamps are stored in descending order. Thus a request for the values of all columns in the row com.cnn.www if no timestamp is specified would be: the value of contents:html from time stamp t6, the value of from time stamp t9, the value of from time stamp t8.

For more information about the internals of how Apache HBase stores data, see Section 9.7, “Regions”.

5.3. Namespace

A namespace is a logical grouping of tables analogous to a database in relation database systems. This abstraction lays the groundwork for upcoming multi-tenancy related features:

  • Quota Management (HBASE-8410) - Restrict the amount of resources (ie regions, tables) a namespace can consume.
  • Namespace Security Administration (HBASE-9206) - provide another level of security administration for tenants.
  • Region server groups (HBASE-6721) - A namespace/table can be pinned onto a subset of regionservers thus guaranteeing a course level of isolation.

5.3.1. Namespace management

A namespace can be created, removed or altered. Namespace membership is determined during table creation by specifying a fully-qualified table name of the form:

<table namespace>:<table qualifier>


#Create a namespace
create_namespace 'my_ns'

#create my_table in my_ns namespace
create 'my_ns:my_table', 'fam'

#delete namespace
delete_namespace 'my_ns'

#alter namespace
alter_namespace 'my_ns', {METHOD => 'set', 'PROPERTY_NAME' => 'PROPERTY_VALUE'}

5.3.2. Predefined namespaces

There are two predefined special namespaces:

  • hbase - system namespace, used to contain hbase internal tables
  • default - tables with no explicit specified namespace will automatically fall into this namespace.


#namespace=foo and table qualifier=bar
create 'foo:bar', 'fam'

#namespace=default and table qualifier=bar
create 'bar', 'fam'

5.4. Table

Tables are declared up front at schema definition time.

5.5. Row

Row keys are uninterrpreted bytes. Rows are lexicographically sorted with the lowest order appearing first in a table. The empty byte array is used to denote both the start and end of a tables' namespace.

5.6. Column Family

Columns in Apache HBase are grouped into column families. All column members of a column family have the same prefix. For example, the columns courses:history and courses:math are both members of the courses column family. The colon character (:) delimits the column family from the . The column family prefix must be composed of printable characters. The qualifying tail, the column family qualifier, can be made of any arbitrary bytes. Column families must be declared up front at schema definition time whereas columns do not need to be defined at schema time but can be conjured on the fly while the table is up an running.

Physically, all column family members are stored together on the filesystem. Because tunings and storage specifications are done at the column family level, it is advised that all column family members have the same general access pattern and size characteristics.

5.7. Cells

A {row, column, version} tuple exactly specifies a cell in HBase. Cell content is uninterrpreted bytes

5.8. Data Model Operations

The four primary data model operations are Get, Put, Scan, and Delete. Operations are applied via HTable instances.

5.8.1. Get

Get returns attributes for a specified row. Gets are executed via HTable.get.

5.8.2. Put

Put either adds new rows to a table (if the key is new) or can update existing rows (if the key already exists). Puts are executed via HTable.put (writeBuffer) or HTable.batch (non-writeBuffer).

5.8.3. Scans

Scan allow iteration over multiple rows for specified attributes.

The following is an example of a on an HTable table instance. Assume that a table is populated with rows with keys "row1", "row2", "row3", and then another set of rows with the keys "abc1", "abc2", and "abc3". The following example shows how startRow and stopRow can be applied to a Scan instance to return the rows beginning with "row".

public static final byte[] CF = "cf".getBytes();
public static final byte[] ATTR = "attr".getBytes();

HTable htable = ...      // instantiate HTable

Scan scan = new Scan();
scan.addColumn(CF, ATTR);
scan.setStartRow(Bytes.toBytes("row")); // start key is inclusive
scan.setStopRow(Bytes.toBytes("rox"));  // stop key is exclusive
ResultScanner rs = htable.getScanner(scan);
try {
  for (Result r =; r != null; r = {
  // process result...
} finally {
  rs.close();  // always close the ResultScanner!

Note that generally the easiest way to specify a specific stop point for a scan is by using the InclusiveStopFilter class.

5.8.4. Delete

Delete removes a row from a table. Deletes are executed via HTable.delete.

HBase does not modify data in place, and so deletes are handled by creating new markers called tombstones. These tombstones, along with the dead values, are cleaned up on major compactions.

See Section, “Delete” for more information on deleting versions of columns, and see Section, “Compaction” for more information on compactions.

5.9. Versions

A {row, column, version} tuple exactly specifies a cell in HBase. It's possible to have an unbounded number of cells where the row and column are the same but the cell address differs only in its version dimension.

While rows and column keys are expressed as bytes, the version is specified using a long integer. Typically this long contains time instances such as those returned by java.util.Date.getTime() or System.currentTimeMillis(), that is: the difference, measured in milliseconds, between the current time and midnight, January 1, 1970 UTC.

The HBase version dimension is stored in decreasing order, so that when reading from a store file, the most recent values are found first.

There is a lot of confusion over the semantics of cell versions, in HBase. In particular, a couple questions that often come up are:

  • If multiple writes to a cell have the same version, are all versions maintained or just the last?[15]

  • Is it OK to write cells in a non-increasing version order?[16]

Below we describe how the version dimension in HBase currently works[17].

5.9.1. Versions and HBase Operations

In this section we look at the behavior of the version dimension for each of the core HBase operations. Get/Scan

Gets are implemented on top of Scans. The below discussion of Get applies equally to Scans.

By default, i.e. if you specify no explicit version, when doing a get, the cell whose version has the largest value is returned (which may or may not be the latest one written, see later). The default behavior can be modified in the following ways:

  • to return more than one version, see Get.setMaxVersions()

  • to return versions other than the latest, see Get.setTimeRange()

    To retrieve the latest version that is less than or equal to a given value, thus giving the 'latest' state of the record at a certain point in time, just use a range from 0 to the desired version and set the max versions to 1. Default Get Example

The following Get will only retrieve the current version of the row

public static final byte[] CF = "cf".getBytes();
public static final byte[] ATTR = "attr".getBytes();
Get get = new Get(Bytes.toBytes("row1"));
Result r = htable.get(get);
byte[] b = r.getValue(CF, ATTR);  // returns current version of value Versioned Get Example

The following Get will return the last 3 versions of the row.

public static final byte[] CF = "cf".getBytes();
public static final byte[] ATTR = "attr".getBytes();
Get get = new Get(Bytes.toBytes("row1"));
get.setMaxVersions(3);  // will return last 3 versions of row
Result r = htable.get(get);
byte[] b = r.getValue(CF, ATTR);  // returns current version of value
List<KeyValue> kv = r.getColumn(CF, ATTR);  // returns all versions of this column Put

Doing a put always creates a new version of a cell, at a certain timestamp. By default the system uses the server's currentTimeMillis, but you can specify the version (= the long integer) yourself, on a per-column level. This means you could assign a time in the past or the future, or use the long value for non-time purposes.

To overwrite an existing value, do a put at exactly the same row, column, and version as that of the cell you would overshadow. Implicit Version Example

The following Put will be implicitly versioned by HBase with the current time.

public static final byte[] CF = "cf".getBytes();
public static final byte[] ATTR = "attr".getBytes();
Put put = new Put(Bytes.toBytes(row));
put.add(CF, ATTR, Bytes.toBytes( data));
htable.put(put); Explicit Version Example

The following Put has the version timestamp explicitly set.

public static final byte[] CF = "cf".getBytes();
public static final byte[] ATTR = "attr".getBytes();
Put put = new Put( Bytes.toBytes(row));
long explicitTimeInMs = 555;  // just an example
put.add(CF, ATTR, explicitTimeInMs, Bytes.toBytes(data));

Caution: the version timestamp is internally by HBase for things like time-to-live calculations. It's usually best to avoid setting this timestamp yourself. Prefer using a separate timestamp attribute of the row, or have the timestamp a part of the rowkey, or both. Delete

There are three different types of internal delete markers [18]:

  • Delete: for a specific version of a column.

  • Delete column: for all versions of a column.

  • Delete family: for all columns of a particular ColumnFamily

When deleting an entire row, HBase will internally create a tombstone for each ColumnFamily (i.e., not each individual column).

Deletes work by creating tombstone markers. For example, let's suppose we want to delete a row. For this you can specify a version, or else by default the currentTimeMillis is used. What this means is delete all cells where the version is less than or equal to this version. HBase never modifies data in place, so for example a delete will not immediately delete (or mark as deleted) the entries in the storage file that correspond to the delete condition. Rather, a so-called tombstone is written, which will mask the deleted values[19]. If the version you specified when deleting a row is larger than the version of any value in the row, then you can consider the complete row to be deleted.

For an informative discussion on how deletes and versioning interact, see the thread Put w/ timestamp -> Deleteall -> Put w/ timestamp fails up on the user mailing list.

Also see Section, “KeyValue” for more information on the internal KeyValue format.

5.9.2. Current Limitations Deletes mask Puts

Deletes mask puts, even puts that happened after the delete was entered[20]. Remember that a delete writes a tombstone, which only disappears after then next major compaction has run. Suppose you do a delete of everything <= T. After this you do a new put with a timestamp <= T. This put, even if it happened after the delete, will be masked by the delete tombstone. Performing the put will not fail, but when you do a get you will notice the put did have no effect. It will start working again after the major compaction has run. These issues should not be a problem if you use always-increasing versions for new puts to a row. But they can occur even if you do not care about time: just do delete and put immediately after each other, and there is some chance they happen within the same millisecond. Major compactions change query results

...create three cell versions at t1, t2 and t3, with a maximum-versions setting of 2. So when getting all versions, only the values at t2 and t3 will be returned. But if you delete the version at t2 or t3, the one at t1 will appear again. Obviously, once a major compaction has run, such behavior will not be the case anymore...[21]

5.10. Sort Order

All data model operations HBase return data in sorted order. First by row, then by ColumnFamily, followed by column qualifier, and finally timestamp (sorted in reverse, so newest records are returned first).

5.11. Column Metadata

There is no store of column metadata outside of the internal KeyValue instances for a ColumnFamily. Thus, while HBase can support not only a wide number of columns per row, but a heterogenous set of columns between rows as well, it is your responsibility to keep track of the column names.

The only way to get a complete set of columns that exist for a ColumnFamily is to process all the rows. For more information about how HBase stores data internally, see Section, “KeyValue”.

5.12. Joins

Whether HBase supports joins is a common question on the dist-list, and there is a simple answer: it doesn't, at not least in the way that RDBMS' support them (e.g., with equi-joins or outer-joins in SQL). As has been illustrated in this chapter, the read data model operations in HBase are Get and Scan.

However, that doesn't mean that equivalent join functionality can't be supported in your application, but you have to do it yourself. The two primary strategies are either denormalizing the data upon writing to HBase, or to have lookup tables and do the join between HBase tables in your application or MapReduce code (and as RDBMS' demonstrate, there are several strategies for this depending on the size of the tables, e.g., nested loops vs. hash-joins). So which is the best approach? It depends on what you are trying to do, and as such there isn't a single answer that works for every use case.

5.13. ACID

<pre>See ACID Semantics. Lars Hofhansl has also written a note on ACID in HBase.</pre>

[15] Currently, only the last written is fetchable.

[16] Yes

[17] See HBASE-2406 for discussion of HBase versions. Bending time in HBase makes for a good read on the version, or time, dimension in HBase. It has more detail on versioning than is provided here. As of this writing, the limiitation Overwriting values at existing timestamps mentioned in the article no longer holds in HBase. This section is basically a synopsis of this article by Bruno Dumon.

[18] See Lars Hofhansl's blog for discussion of his attempt adding another, Scanning in HBase: Prefix Delete Marker

[19] When HBase does a major compaction, the tombstones are processed to actually remove the dead values, together with the tombstones themselves.

[21] See Garbage Collection in Bending time in HBase

Chapter 6. HBase and Schema Design

A good general introduction on the strength and weaknesses modelling on the various non-rdbms datastores is Ian Varley's Master thesis, No Relation: The Mixed Blessings of Non-Relational Databases. Recommended. Also, read Section, “KeyValue” for how HBase stores data internally, and the section on Section 6.11, “Schema Design Case Studies”.

6.1.  Schema Creation

HBase schemas can be created or updated with Chapter 4, The Apache HBase Shell or by using HBaseAdmin in the Java API.

Tables must be disabled when making ColumnFamily modifications, for example..

Configuration config = HBaseConfiguration.create();
HBaseAdmin admin = new HBaseAdmin(conf);
String table = "myTable";


HColumnDescriptor cf1 = ...;
admin.addColumn(table, cf1);      // adding new ColumnFamily
HColumnDescriptor cf2 = ...;
admin.modifyColumn(table, cf2);    // modifying existing ColumnFamily


See Section 2.3.4, “Client configuration and dependencies connecting to an HBase cluster” for more information about configuring client connections.

Note: online schema changes are supported in the 0.92.x codebase, but the 0.90.x codebase requires the table to be disabled.

6.1.1. Schema Updates

When changes are made to either Tables or ColumnFamilies (e.g., region size, block size), these changes take effect the next time there is a major compaction and the StoreFiles get re-written.

See Section 9.7.6, “Store” for more information on StoreFiles.

6.2.  On the number of column families

HBase currently does not do well with anything above two or three column families so keep the number of column families in your schema low. Currently, flushing and compactions are done on a per Region basis so if one column family is carrying the bulk of the data bringing on flushes, the adjacent families will also be flushed though the amount of data they carry is small. When many column families the flushing and compaction interaction can make for a bunch of needless i/o loading (To be addressed by changing flushing and compaction to work on a per column family basis). For more information on compactions, see Section, “Compaction”.

Try to make do with one column family if you can in your schemas. Only introduce a second and third column family in the case where data access is usually column scoped; i.e. you query one column family or the other but usually not both at the one time.

6.2.1. Cardinality of ColumnFamilies

Where multiple ColumnFamilies exist in a single table, be aware of the cardinality (i.e., number of rows). If ColumnFamilyA has 1 million rows and ColumnFamilyB has 1 billion rows, ColumnFamilyA's data will likely be spread across many, many regions (and RegionServers). This makes mass scans for ColumnFamilyA less efficient.

6.3. Rowkey Design

6.3.1.  Monotonically Increasing Row Keys/Timeseries Data

In the HBase chapter of Tom White's book Hadoop: The Definitive Guide (O'Reilly) there is a an optimization note on watching out for a phenomenon where an import process walks in lock-step with all clients in concert pounding one of the table's regions (and thus, a single node), then moving onto the next region, etc. With monotonically increasing row-keys (i.e., using a timestamp), this will happen. See this comic by IKai Lan on why monotonically increasing row keys are problematic in BigTable-like datastores: monotonically increasing values are bad. The pile-up on a single region brought on by monotonically increasing keys can be mitigated by randomizing the input records to not be in sorted order, but in general it's best to avoid using a timestamp or a sequence (e.g. 1, 2, 3) as the row-key.

If you do need to upload time series data into HBase, you should study OpenTSDB as a successful example. It has a page describing the schema it uses in HBase. The key format in OpenTSDB is effectively [metric_type][event_timestamp], which would appear at first glance to contradict the previous advice about not using a timestamp as the key. However, the difference is that the timestamp is not in the lead position of the key, and the design assumption is that there are dozens or hundreds (or more) of different metric types. Thus, even with a continual stream of input data with a mix of metric types, the Puts are distributed across various points of regions in the table.

See Section 6.11, “Schema Design Case Studies” for some rowkey design examples.

6.3.2. Try to minimize row and column sizes

Or why are my StoreFile indices large?

In HBase, values are always freighted with their coordinates; as a cell value passes through the system, it'll be accompanied by its row, column name, and timestamp - always. If your rows and column names are large, especially compared to the size of the cell value, then you may run up against some interesting scenarios. One such is the case described by Marc Limotte at the tail of HBASE-3551 (recommended!). Therein, the indices that are kept on HBase storefiles (Section, “StoreFile (HFile)”) to facilitate random access may end up occupyng large chunks of the HBase allotted RAM because the cell value coordinates are large. Mark in the above cited comment suggests upping the block size so entries in the store file index happen at a larger interval or modify the table schema so it makes for smaller rows and column names. Compression will also make for larger indices. See the thread a question storefileIndexSize up on the user mailing list.

Most of the time small inefficiencies don't matter all that much. Unfortunately, this is a case where they do. Whatever patterns are selected for ColumnFamilies, attributes, and rowkeys they could be repeated several billion times in your data.

See Section, “KeyValue” for more information on HBase stores data internally to see why this is important. Column Families

Try to keep the ColumnFamily names as small as possible, preferably one character (e.g. "d" for data/default).

See Section, “KeyValue” for more information on HBase stores data internally to see why this is important. Attributes

Although verbose attribute names (e.g., "myVeryImportantAttribute") are easier to read, prefer shorter attribute names (e.g., "via") to store in HBase.

See Section, “KeyValue” for more information on HBase stores data internally to see why this is important. Rowkey Length

Keep them as short as is reasonable such that they can still be useful for required data access (e.g., Get vs. Scan). A short key that is useless for data access is not better than a longer key with better get/scan properties. Expect tradeoffs when designing rowkeys. Byte Patterns

A long is 8 bytes. You can store an unsigned number up to 18,446,744,073,709,551,615 in those eight bytes. If you stored this number as a String -- presuming a byte per character -- you need nearly 3x the bytes.

Not convinced? Below is some sample code that you can run on your own.

// long
long l = 1234567890L;
byte[] lb = Bytes.toBytes(l);
System.out.println("long bytes length: " + lb.length);   // returns 8

String s = "" + l;
byte[] sb = Bytes.toBytes(s);
System.out.println("long as string length: " + sb.length);    // returns 10

// hash
MessageDigest md = MessageDigest.getInstance("MD5");
byte[] digest = md.digest(Bytes.toBytes(s));
System.out.println("md5 digest bytes length: " + digest.length);    // returns 16

String sDigest = new String(digest);
byte[] sbDigest = Bytes.toBytes(sDigest);
System.out.println("md5 digest as string length: " + sbDigest.length);    // returns 26

Unfortunately, using a binary representation of a type will make your data harder to read outside of your code. For example, this is what you will see in the shell when you increment a value:

hbase(main):001:0> incr 't', 'r', 'f:q', 1

hbase(main):002:0> get 't', 'r'
COLUMN                                        CELL
 f:q                                          timestamp=1369163040570, value=\x00\x00\x00\x00\x00\x00\x00\x01
1 row(s) in 0.0310 seconds

The shell makes a best effort to print a string, and it this case it decided to just print the hex. The same will happen to your row keys inside the region names. It can be okay if you know what's being stored, but it might also be unreadable if arbitrary data can be put in the same cells. This is the main trade-off.

6.3.3. Reverse Timestamps

A common problem in database processing is quickly finding the most recent version of a value. A technique using reverse timestamps as a part of the key can help greatly with a special case of this problem. Also found in the HBase chapter of Tom White's book Hadoop: The Definitive Guide (O'Reilly), the technique involves appending (Long.MAX_VALUE - timestamp) to the end of any key, e.g., [key][reverse_timestamp].

The most recent value for [key] in a table can be found by performing a Scan for [key] and obtaining the first record. Since HBase keys are in sorted order, this key sorts before any older row-keys for [key] and thus is first.

This technique would be used instead of using Section 6.4, “ Number of Versions ” where the intent is to hold onto all versions "forever" (or a very long time) and at the same time quickly obtain access to any other version by using the same Scan technique.

6.3.4. Rowkeys and ColumnFamilies

Rowkeys are scoped to ColumnFamilies. Thus, the same rowkey could exist in each ColumnFamily that exists in a table without collision.

6.3.5. Immutability of Rowkeys

Rowkeys cannot be changed. The only way they can be "changed" in a table is if the row is deleted and then re-inserted. This is a fairly common question on the HBase dist-list so it pays to get the rowkeys right the first time (and/or before you've inserted a lot of data).

6.3.6. Relationship Between RowKeys and Region Splits

If you pre-split your table, it is critical to understand how your rowkey will be distributed across the region boundaries. As an example of why this is important, consider the example of using displayable hex characters as the lead position of the key (e.g., ""0000000000000000" to "ffffffffffffffff"). Running those key ranges through Bytes.split (which is the split strategy used when creating regions in HBaseAdmin.createTable(byte[] startKey, byte[] endKey, numRegions) for 10 regions will generate the following splits...

48 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48                                // 0
54 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 -10                 // 6
61 -67 -67 -67 -67 -67 -67 -67 -67 -67 -67 -67 -67 -67 -67 -68                 // =
68 -124 -124 -124 -124 -124 -124 -124 -124 -124 -124 -124 -124 -124 -124 -126  // D
75 75 75 75 75 75 75 75 75 75 75 75 75 75 75 72                                // K
82 18 18 18 18 18 18 18 18 18 18 18 18 18 18 14                                // R
88 -40 -40 -40 -40 -40 -40 -40 -40 -40 -40 -40 -40 -40 -40 -44                 // X
95 -97 -97 -97 -97 -97 -97 -97 -97 -97 -97 -97 -97 -97 -97 -102                // _
102 102 102 102 102 102 102 102 102 102 102 102 102 102 102 102                // f

... (note: the lead byte is listed to the right as a comment.) Given that the first split is a '0' and the last split is an 'f', everything is great, right? Not so fast.

The problem is that all the data is going to pile up in the first 2 regions and the last region thus creating a "lumpy" (and possibly "hot") region problem. To understand why, refer to an ASCII Table. '0' is byte 48, and 'f' is byte 102, but there is a huge gap in byte values (bytes 58 to 96) that will never appear in this keyspace because the only values are [0-9] and [a-f]. Thus, the middle regions regions will never be used. To make pre-spliting work with this example keyspace, a custom definition of splits (i.e., and not relying on the built-in split method) is required.

Lesson #1: Pre-splitting tables is generally a best practice, but you need to pre-split them in such a way that all the regions are accessible in the keyspace. While this example demonstrated the problem with a hex-key keyspace, the same problem can happen with any keyspace. Know your data.

Lesson #2: While generally not advisable, using hex-keys (and more generally, displayable data) can still work with pre-split tables as long as all the created regions are accessible in the keyspace.

To conclude this example, the following is an example of how appropriate splits can be pre-created for hex-keys:.

public static boolean createTable(HBaseAdmin admin, HTableDescriptor table, byte[][] splits)
throws IOException {
  try {
    admin.createTable( table, splits );
    return true;
  } catch (TableExistsException e) {"table " + table.getNameAsString() + " already exists");
    // the table already exists...
    return false;

public static byte[][] getHexSplits(String startKey, String endKey, int numRegions) {
  byte[][] splits = new byte[numRegions-1][];
  BigInteger lowestKey = new BigInteger(startKey, 16);
  BigInteger highestKey = new BigInteger(endKey, 16);
  BigInteger range = highestKey.subtract(lowestKey);
  BigInteger regionIncrement = range.divide(BigInteger.valueOf(numRegions));
  lowestKey = lowestKey.add(regionIncrement);
  for(int i=0; i < numRegions-1;i++) {
    BigInteger key = lowestKey.add(regionIncrement.multiply(BigInteger.valueOf(i)));
    byte[] b = String.format("%016x", key).getBytes();
    splits[i] = b;
  return splits;

6.4.  Number of Versions

6.4.1. Maximum Number of Versions

The maximum number of row versions to store is configured per column family via HColumnDescriptor. The default for max versions is 3. This is an important parameter because as described in Chapter 5, Data Model section HBase does not overwrite row values, but rather stores different values per row by time (and qualifier). Excess versions are removed during major compactions. The number of max versions may need to be increased or decreased depending on application needs.

It is not recommended setting the number of max versions to an exceedingly high level (e.g., hundreds or more) unless those old values are very dear to you because this will greatly increase StoreFile size.

6.4.2.  Minimum Number of Versions

Like maximum number of row versions, the minimum number of row versions to keep is configured per column family via HColumnDescriptor. The default for min versions is 0, which means the feature is disabled. The minimum number of row versions parameter is used together with the time-to-live parameter and can be combined with the number of row versions parameter to allow configurations such as "keep the last T minutes worth of data, at most N versions, but keep at least M versions around" (where M is the value for minimum number of row versions, M<N). This parameter should only be set when time-to-live is enabled for a column family and must be less than the number of row versions.

6.5.  Supported Datatypes

HBase supports a "bytes-in/bytes-out" interface via Put and Result, so anything that can be converted to an array of bytes can be stored as a value. Input could be strings, numbers, complex objects, or even images as long as they can rendered as bytes.

There are practical limits to the size of values (e.g., storing 10-50MB objects in HBase would probably be too much to ask); search the mailling list for conversations on this topic. All rows in HBase conform to the Chapter 5, Data Model, and that includes versioning. Take that into consideration when making your design, as well as block size for the ColumnFamily.

6.5.1. Counters

One supported datatype that deserves special mention are "counters" (i.e., the ability to do atomic increments of numbers). See Increment in HTable.

Synchronization on counters are done on the RegionServer, not in the client.

6.6. Joins

If you have multiple tables, don't forget to factor in the potential for Section 5.12, “Joins” into the schema design.

6.7. Time To Live (TTL)

ColumnFamilies can set a TTL length in seconds, and HBase will automatically delete rows once the expiration time is reached. This applies to all versions of a row - even the current one. The TTL time encoded in the HBase for the row is specified in UTC.

See HColumnDescriptor for more information.

6.8.  Keeping Deleted Cells

ColumnFamilies can optionally keep deleted cells. That means deleted cells can still be retrieved with Get or Scan operations, as long these operations have a time range specified that ends before the timestamp of any delete that would affect the cells. This allows for point in time queries even in the presence of deletes.

Deleted cells are still subject to TTL and there will never be more than "maximum number of versions" deleted cells. A new "raw" scan options returns all deleted rows and the delete markers.

See HColumnDescriptor for more information.

6.9.  Secondary Indexes and Alternate Query Paths

This section could also be titled "what if my table rowkey looks like this but I also want to query my table like that." A common example on the dist-list is where a row-key is of the format "user-timestamp" but there are reporting requirements on activity across users for certain time ranges. Thus, selecting by user is easy because it is in the lead position of the key, but time is not.

There is no single answer on the best way to handle this because it depends on...

  • Number of users
  • Data size and data arrival rate
  • Flexibility of reporting requirements (e.g., completely ad-hoc date selection vs. pre-configured ranges)
  • Desired execution speed of query (e.g., 90 seconds may be reasonable to some for an ad-hoc report, whereas it may be too long for others)

... and solutions are also influenced by the size of the cluster and how much processing power you have to throw at the solution. Common techniques are in sub-sections below. This is a comprehensive, but not exhaustive, list of approaches.

It should not be a surprise that secondary indexes require additional cluster space and processing. This is precisely what happens in an RDBMS because the act of creating an alternate index requires both space and processing cycles to update. RBDMS products are more advanced in this regard to handle alternative index management out of the box. However, HBase scales better at larger data volumes, so this is a feature trade-off.

Pay attention to Chapter 12, Apache HBase Performance Tuning when implementing any of these approaches.

Additionally, see the David Butler response in this dist-list thread HBase, mail # user - Stargate+hbase

6.9.1.  Filter Query

Depending on the case, it may be appropriate to use Section 9.4, “Client Request Filters”. In this case, no secondary index is created. However, don't try a full-scan on a large table like this from an application (i.e., single-threaded client).

6.9.2.  Periodic-Update Secondary Index

A secondary index could be created in an other table which is periodically updated via a MapReduce job. The job could be executed intra-day, but depending on load-strategy it could still potentially be out of sync with the main data table.

See Section 7.2.2, “HBase MapReduce Read/Write Example” for more information.

6.9.3.  Dual-Write Secondary Index

Another strategy is to build the secondary index while publishing data to the cluster (e.g., write to data table, write to index table). If this is approach is taken after a data table already exists, then bootstrapping will be needed for the secondary index with a MapReduce job (see Section 6.9.2, “ Periodic-Update Secondary Index ”).

6.9.4.  Summary Tables

Where time-ranges are very wide (e.g., year-long report) and where the data is voluminous, summary tables are a common approach. These would be generated with MapReduce jobs into another table.

See Section 7.2.4, “HBase MapReduce Summary to HBase Example” for more information.

6.9.5.  Coprocessor Secondary Index

Coprocessors act like RDBMS triggers. These were added in 0.92. For more information, see Section 9.6.3, “Coprocessors”

6.10. Constraints

HBase currently supports 'constraints' in traditional (SQL) database parlance. The advised usage for Constraints is in enforcing business rules for attributes in the table (eg. make sure values are in the range 1-10). Constraints could also be used to enforce referential integrity, but this is strongly discouraged as it will dramatically decrease the write throughput of the tables where integrity checking is enabled. Extensive documentation on using Constraints can be found at: Constraint since version 0.94.

6.11. Schema Design Case Studies

The following will describe some typical data ingestion use-cases with HBase, and how the rowkey design and construction can be approached. Note: this is just an illustration of potential approaches, not an exhaustive list. Know your data, and know your processing requirements.

It is highly recommended that you read the rest of the Chapter 6, HBase and Schema Design first, before reading these case studies.

Thee following case studies are described:

  • Log Data / Timeseries Data
  • Log Data / Timeseries on Steroids
  • Customer/Order
  • Tall/Wide/Middle Schema Design
  • List Data

6.11.1. Case Study - Log Data and Timeseries Data

Assume that the following data elements are being collected.

  • Hostname
  • Timestamp
  • Log event
  • Value/message

We can store them in an HBase table called LOG_DATA, but what will the rowkey be? From these attributes the rowkey will be some combination of hostname, timestamp, and log-event - but what specifically? Timestamp In The Rowkey Lead Position

The rowkey [timestamp][hostname][log-event] suffers from the monotonically increasing rowkey problem described in Section 6.3.1, “ Monotonically Increasing Row Keys/Timeseries Data ”.

There is another pattern frequently mentioned in the dist-lists about “bucketing” timestamps, by performing a mod operation on the timestamp. If time-oriented scans are important, this could be a useful approach. Attention must be paid to the number of buckets, because this will require the same number of scans to return results.

long bucket = timestamp % numBuckets;

… to construct:


As stated above, to select data for a particular timerange, a Scan will need to be performed for each bucket. 100 buckets, for example, will provide a wide distribution in the keyspace but it will require 100 Scans to obtain data for a single timestamp, so there are trade-offs. Host In The Rowkey Lead Position

The rowkey [hostname][log-event][timestamp] is a candidate if there is a large-ish number of hosts to spread the writes and reads across the keyspace. This approach would be useful if scanning by hostname was a priority. Timestamp, or Reverse Timestamp?

If the most important access path is to pull most recent events, then storing the timestamps as reverse-timestamps (e.g., timestamp = Long.MAX_VALUE – timestamp) will create the property of being able to do a Scan on [hostname][log-event] to obtain the quickly obtain the most recently captured events.

Neither approach is wrong, it just depends on what is most appropriate for the situation. Variangle Length or Fixed Length Rowkeys?

It is critical to remember that rowkeys are stamped on every column in HBase. If the hostname is “a” and the event type is “e1” then the resulting rowkey would be quite small. However, what if the ingested hostname is “” and the event type is “com.package1.subpackage2.subsubpackage3.ImportantService”?

It might make sense to use some substitution in the rowkey. There are at least two approaches: hashed and numeric. In the Hostname In The Rowkey Lead Position example, it might look like this:

Composite Rowkey With Hashes:

  • [MD5 hash of hostname] = 16 bytes
  • [MD5 hash of event-type] = 16 bytes
  • [timestamp] = 8 bytes

Composite Rowkey With Numeric Substitution:

For this approach another lookup table would be needed in addition to LOG_DATA, called LOG_TYPES. The rowkey of LOG_TYPES would be:

  • [type] (e.g., byte indicating hostname vs. event-type)
  • [bytes] variable length bytes for raw hostname or event-type.

A column for this rowkey could be a long with an assigned number, which could be obtained by using an HBase counter.

So the resulting composite rowkey would be:

  • [substituted long for hostname] = 8 bytes
  • [substituted long for event type] = 8 bytes
  • [timestamp] = 8 bytes

In either the Hash or Numeric substitution approach, the raw values for hostname and event-type can be stored as columns.

6.11.2. Case Study - Log Data and Timeseries Data on Steroids

This effectively is the OpenTSDB approach. What OpenTSDB does is re-write data and pack rows into columns for certain time-periods. For a detailed explanation, see:, and Lessons Learned from OpenTSDB from HBaseCon2012.

But this is how the general concept works: data is ingested, for example, in this manner…


… with separate rowkeys for each detailed event, but is re-written like this…


… and each of the above events are converted into columns stored with a time-offset relative to the beginning timerange (e.g., every 5 minutes). This is obviously a very advanced processing technique, but HBase makes this possible.

6.11.3. Case Study - Customer/Order

Assume that HBase is used to store customer and order information. There are two core record-types being ingested: a Customer record type, and Order record type.

The Customer record type would include all the things that you’d typically expect:

  • Customer number
  • Customer name
  • Address (e.g., city, state, zip)
  • Phone numbers, etc.

The Order record type would include things like:

Assuming that the combination of customer number and sales order uniquely identify an order, these two attributes will compose the rowkey, and specifically a composite key such as:

[customer number][order number]

… for a ORDER table. However, there are more design decisions to make: are the raw values the best choices for rowkeys?

The same design questions in the Log Data use-case confront us here. What is the keyspace of the customer number, and what is the format (e.g., numeric? alphanumeric?) As it is advantageous to use fixed-length keys in HBase, as well as keys that can support a reasonable spread in the keyspace, similar options appear:

Composite Rowkey With Hashes:

  • [MD5 of customer number] = 16 bytes
  • [MD5 of order number] = 16 bytes

Composite Numeric/Hash Combo Rowkey:

  • [substituted long for customer number] = 8 bytes
  • [MD5 of order number] = 16 bytes Single Table? Multiple Tables?

A traditional design approach would have separate tables for CUSTOMER and SALES. Another option is to pack multiple record types into a single table (e.g., CUSTOMER++).

Customer Record Type Rowkey:

  • [customer-id]
  • [type] = type indicating ‘1’ for customer record type

Order Record Type Rowkey:

  • [customer-id]
  • [type] = type indicating ‘2’ for order record type
  • [order]

The advantage of this particular CUSTOMER++ approach is that organizes many different record-types by customer-id (e.g., a single scan could get you everything about that customer). The disadvantage is that it’s not as easy to scan for a particular record-type. Order Object Design

Now we need to address how to model the Order object. Assume that the class structure is as follows:

     ShippingLocation     (an Order can have multiple ShippingLocations)
          LineItem               (a ShippingLocation can have multiple LineItems)

... there are multiple options on storing this data. Completely Normalized

With this approach, there would be separate tables for ORDER, SHIPPING_LOCATION, and LINE_ITEM.

The ORDER table's rowkey was described above: Section 6.11.3, “Case Study - Customer/Order”

The SHIPPING_LOCATION's composite rowkey would be something like this:

  • [order-rowkey]
  • [shipping location number] (e.g., 1st location, 2nd, etc.)

The LINE_ITEM table's composite rowkey would be something like this:

  • [order-rowkey]
  • [shipping location number] (e.g., 1st location, 2nd, etc.)
  • [line item number] (e.g., 1st lineitem, 2nd, etc.)

Such a normalized model is likely to be the approach with an RDBMS, but that's not your only option with HBase. The cons of such an approach is that to retrieve information about any Order, you will need:

  • Get on the ORDER table for the Order
  • Scan on the SHIPPING_LOCATION table for that order to get the ShippingLocation instances
  • Scan on the LINE_ITEM for each ShippingLocation

... granted, this is what an RDBMS would do under the covers anyway, but since there are no joins in HBase you're just more aware of this fact. Single Table With Record Types

With this approach, there would exist a single table ORDER that would contain

The Order rowkey was described above: Section 6.11.3, “Case Study - Customer/Order”

  • [order-rowkey]
  • [ORDER record type]

The ShippingLocation composite rowkey would be something like this:

  • [order-rowkey]
  • [SHIPPING record type]
  • [shipping location number] (e.g., 1st location, 2nd, etc.)

The LineItem composite rowkey would be something like this:

  • [order-rowkey]
  • [LINE record type]
  • [shipping location number] (e.g., 1st location, 2nd, etc.)
  • [line item number] (e.g., 1st lineitem, 2nd, etc.) Denormalized

A variant of the Single Table With Record Types approach is to denormalize and flatten some of the object hierarchy, such as collapsing the ShippingLocation attributes onto each LineItem instance.

The LineItem composite rowkey would be something like this:

  • [order-rowkey]
  • [LINE record type]
  • [line item number] (e.g., 1st lineitem, 2nd, etc. - care must be taken that there are unique across the entire order)

... and the LineItem columns would be something like this:

  • itemNumber
  • quantity
  • price
  • shipToLine1 (denormalized from ShippingLocation)
  • shipToLine2 (denormalized from ShippingLocation)
  • shipToCity (denormalized from ShippingLocation)
  • shipToState (denormalized from ShippingLocation)
  • shipToZip (denormalized from ShippingLocation)

The pros of this approach include a less complex object heirarchy, but one of the cons is that updating gets more complicated in case any of this information changes. Object BLOB

With this approach, the entire Order object graph is treated, in one way or another, as a BLOB. For example, the ORDER table's rowkey was described above: Section 6.11.3, “Case Study - Customer/Order”, and a single column called "order" would contain an object that could be deserialized that contained a container Order, ShippingLocations, and LineItems.

There are many options here: JSON, XML, Java Serialization, Avro, Hadoop Writables, etc. All of them are variants of the same approach: encode the object graph to a byte-array. Care should be taken with this approach to ensure backward compatibilty in case the object model changes such that older persisted structures can still be read back out of HBase.

Pros are being able to manage complex object graphs with minimal I/O (e.g., a single HBase Get per Order in this example), but the cons include the aforementioned warning about backward compatiblity of serialization, language dependencies of serialization (e.g., Java Serialization only works with Java clients), the fact that you have to deserialize the entire object to get any piece of information inside the BLOB, and the difficulty in getting frameworks like Hive to work with custom objects like this.

6.11.4. Case Study - "Tall/Wide/Middle" Schema Design Smackdown

This section will describe additional schema design questions that appear on the dist-list, specifically about tall and wide tables. These are general guidelines and not laws - each application must consider its own needs. Rows vs. Versions

A common question is whether one should prefer rows or HBase's built-in-versioning. The context is typically where there are "a lot" of versions of a row to be retained (e.g., where it is significantly above the HBase default of 3 max versions). The rows-approach would require storing a timstamp in some portion of the rowkey so that they would not overwite with each successive update.

Preference: Rows (generally speaking). Rows vs. Columns

Another common question is whether one should prefer rows or columns. The context is typically in extreme cases of wide tables, such as having 1 row with 1 million attributes, or 1 million rows with 1 columns apiece.

Preference: Rows (generally speaking). To be clear, this guideline is in the context is in extremely wide cases, not in the standard use-case where one needs to store a few dozen or hundred columns. But there is also a middle path between these two options, and that is "Rows as Columns." Rows as Columns

The middle path between Rows vs. Columns is packing data that would be a separate row into columns, for certain rows. OpenTSDB is the best example of this case where a single row represents a defined time-range, and then discrete events are treated as columns. This approach is often more complex, and may require the additional complexity of re-writing your data, but has the advantage of being I/O efficient. For an overview of this approach, see ???.

6.11.5. Case Study - List Data

The following is an exchange from the user dist-list regarding a fairly common question: how to handle per-user list data in Apache HBase.

*** QUESTION ***

We're looking at how to store a large amount of (per-user) list data in HBase, and we were trying to figure out what kind of access pattern made the most sense. One option is store the majority of the data in a key, so we could have something like:

<FixedWidthUserName><FixedWidthValueId1>:"" (no value)
<FixedWidthUserName><FixedWidthValueId2>:"" (no value)
<FixedWidthUserName><FixedWidthValueId3>:"" (no value)
The other option we had was to do this entirely using:

where each row would contain multiple values. So in one case reading the first thirty values would be:

scan { STARTROW => 'FixedWidthUsername' LIMIT => 30}
And in the second case it would be
get 'FixedWidthUserName\x00\x00\x00\x00'

The general usage pattern would be to read only the first 30 values of these lists, with infrequent access reading deeper into the lists. Some users would have <= 30 total values in these lists, and some users would have millions (i.e. power-law distribution)

The single-value format seems like it would take up more space on HBase, but would offer some improved retrieval / pagination flexibility. Would there be any significant performance advantages to be able to paginate via gets vs paginating with scans?

My initial understanding was that doing a scan should be faster if our paging size is unknown (and caching is set appropriately), but that gets should be faster if we'll always need the same page size. I've ended up hearing different people tell me opposite things about performance. I assume the page sizes would be relatively consistent, so for most use cases we could guarantee that we only wanted one page of data in the fixed-page-length case. I would also assume that we would have infrequent updates, but may have inserts into the middle of these lists (meaning we'd need to update all subsequent rows).

Thanks for help / suggestions / follow-up questions.

*** ANSWER ***

If I understand you correctly, you're ultimately trying to store triples in the form "user, valueid, value", right? E.g., something like:

"user123, firstname, Paul",
"user234, lastname, Smith"

(But the usernames are fixed width, and the valueids are fixed width).

And, your access pattern is along the lines of: "for user X, list the next 30 values, starting with valueid Y". Is that right? And these values should be returned sorted by valueid?

The tl;dr version is that you should probably go with one row per user+value, and not build a complicated intra-row pagination scheme on your own unless you're really sure it is needed.

Your two options mirror a common question people have when designing HBase schemas: should I go "tall" or "wide"? Your first schema is "tall": each row represents one value for one user, and so there are many rows in the table for each user; the row key is user + valueid, and there would be (presumably) a single column qualifier that means "the value". This is great if you want to scan over rows in sorted order by row key (thus my question above, about whether these ids are sorted correctly). You can start a scan at any user+valueid, read the next 30, and be done. What you're giving up is the ability to have transactional guarantees around all the rows for one user, but it doesn't sound like you need that. Doing it this way is generally recommended (see here

Your second option is "wide": you store a bunch of values in one row, using different qualifiers (where the qualifier is the valueid). The simple way to do that would be to just store ALL values for one user in a single row. I'm guessing you jumped to the "paginated" version because you're assuming that storing millions of columns in a single row would be bad for performance, which may or may not be true; as long as you're not trying to do too much in a single request, or do things like scanning over and returning all of the cells in the row, it shouldn't be fundamentally worse. The client has methods that allow you to get specific slices of columns.

Note that neither case fundamentally uses more disk space than the other; you're just "shifting" part of the identifying information for a value either to the left (into the row key, in option one) or to the right (into the column qualifiers in option 2). Under the covers, every key/value still stores the whole row key, and column family name. (If this is a bit confusing, take an hour and watch Lars George's excellent video about understanding HBase schema design:

A manually paginated version has lots more complexities, as you note, like having to keep track of how many things are in each page, re-shuffling if new values are inserted, etc. That seems significantly more complex. It might have some slight speed advantages (or disadvantages!) at extremely high throughput, and the only way to really know that would be to try it out. If you don't have time to build it both ways and compare, my advice would be to start with the simplest option (one row per user+value). Start simple and iterate! :)

6.12. Operational and Performance Configuration Options

See the Performance section Section 12.6, “Schema Design” for more information operational and performance schema design options, such as Bloom Filters, Table-configured regionsizes, compression, and blocksizes.

Chapter 7. HBase and MapReduce

See HBase and MapReduce up in javadocs. Start there. Below is some additional help.

For more information about MapReduce (i.e., the framework in general), see the Hadoop MapReduce Tutorial.

7.1. Map-Task Splitting

7.1.1. The Default HBase MapReduce Splitter

When TableInputFormat is used to source an HBase table in a MapReduce job, its splitter will make a map task for each region of the table. Thus, if there are 100 regions in the table, there will be 100 map-tasks for the job - regardless of how many column families are selected in the Scan.

7.1.2. Custom Splitters

For those interested in implementing custom splitters, see the method getSplits in TableInputFormatBase. That is where the logic for map-task assignment resides.

7.2. HBase MapReduce Examples

7.2.1. HBase MapReduce Read Example

The following is an example of using HBase as a MapReduce source in read-only manner. Specifically, there is a Mapper instance but no Reducer, and nothing is being emitted from the Mapper. There job would be defined as follows...

Configuration config = HBaseConfiguration.create();
Job job = new Job(config, "ExampleRead");
job.setJarByClass(MyReadJob.class);     // class that contains mapper

Scan scan = new Scan();
scan.setCaching(500);        // 1 is the default in Scan, which will be bad for MapReduce jobs
scan.setCacheBlocks(false);  // don't set to true for MR jobs
// set other scan attrs

  tableName,        // input HBase table name
  scan,             // Scan instance to control CF and attribute selection
  MyMapper.class,   // mapper
  null,             // mapper output key
  null,             // mapper output value
job.setOutputFormatClass(NullOutputFormat.class);   // because we aren't emitting anything from mapper

boolean b = job.waitForCompletion(true);
if (!b) {
  throw new IOException("error with job!");

...and the mapper instance would extend TableMapper...

public static class MyMapper extends TableMapper<Text, Text> {

  public void map(ImmutableBytesWritable row, Result value, Context context) throws InterruptedException, IOException {
    // process data for the row from the Result instance.

7.2.2. HBase MapReduce Read/Write Example

The following is an example of using HBase both as a source and as a sink with MapReduce. This example will simply copy data from one table to another.

Configuration config = HBaseConfiguration.create();
Job job = new Job(config,"ExampleReadWrite");
job.setJarByClass(MyReadWriteJob.class);    // class that contains mapper

Scan scan = new Scan();
scan.setCaching(500);        // 1 is the default in Scan, which will be bad for MapReduce jobs
scan.setCacheBlocks(false);  // don't set to true for MR jobs
// set other scan attrs

	sourceTable,      // input table
	scan,	          // Scan instance to control CF and attribute selection
	MyMapper.class,   // mapper class
	null,	          // mapper output key
	null,	          // mapper output value
	targetTable,      // output table
	null,             // reducer class

boolean b = job.waitForCompletion(true);
if (!b) {
    throw new IOException("error with job!");

An explanation is required of what TableMapReduceUtil is doing, especially with the reducer. TableOutputFormat is being used as the outputFormat class, and several parameters are being set on the config (e.g., TableOutputFormat.OUTPUT_TABLE), as well as setting the reducer output key to ImmutableBytesWritable and reducer value to Writable. These could be set by the programmer on the job and conf, but TableMapReduceUtil tries to make things easier.

The following is the example mapper, which will create a Put and matching the input Result and emit it. Note: this is what the CopyTable utility does.

public static class MyMapper extends TableMapper<ImmutableBytesWritable, Put>  {

	public void map(ImmutableBytesWritable row, Result value, Context context) throws IOException, InterruptedException {
		// this example is just copying the data from the source table...
   		context.write(row, resultToPut(row,value));

  	private static Put resultToPut(ImmutableBytesWritable key, Result result) throws IOException {
  		Put put = new Put(key.get());
 		for (KeyValue kv : result.raw()) {
		return put;

There isn't actually a reducer step, so TableOutputFormat takes care of sending the Put to the target table.

This is just an example, developers could choose not to use TableOutputFormat and connect to the target table themselves.

7.2.3. HBase MapReduce Read/Write Example With Multi-Table Output

TODO: example for MultiTableOutputFormat.

7.2.4. HBase MapReduce Summary to HBase Example

The following example uses HBase as a MapReduce source and sink with a summarization step. This example will count the number of distinct instances of a value in a table and write those summarized counts in another table.

Configuration config = HBaseConfiguration.create();
Job job = new Job(config,"ExampleSummary");
job.setJarByClass(MySummaryJob.class);     // class that contains mapper and reducer

Scan scan = new Scan();
scan.setCaching(500);        // 1 is the default in Scan, which will be bad for MapReduce jobs
scan.setCacheBlocks(false);  // don't set to true for MR jobs
// set other scan attrs

	sourceTable,        // input table
	scan,               // Scan instance to control CF and attribute selection
	MyMapper.class,     // mapper class
	Text.class,         // mapper output key
	IntWritable.class,  // mapper output value
	targetTable,        // output table
	MyTableReducer.class,    // reducer class
job.setNumReduceTasks(1);   // at least one, adjust as required

boolean b = job.waitForCompletion(true);
if (!b) {
	throw new IOException("error with job!");

In this example mapper a column with a String-value is chosen as the value to summarize upon. This value is used as the key to emit from the mapper, and an IntWritable represents an instance counter.

public static class MyMapper extends TableMapper<Text, IntWritable>  {
	public static final byte[] CF = "cf".getBytes();
	public static final byte[] ATTR1 = "attr1".getBytes();

	private final IntWritable ONE = new IntWritable(1);
   	private Text text = new Text();

   	public void map(ImmutableBytesWritable row, Result value, Context context) throws IOException, InterruptedException {
        	String val = new String(value.getValue(CF, ATTR1));
          	text.set(val);     // we can only emit Writables...

        	context.write(text, ONE);

In the reducer, the "ones" are counted (just like any other MR example that does this), and then emits a Put.

public static class MyTableReducer extends TableReducer<Text, IntWritable, ImmutableBytesWritable>  {
	public static final byte[] CF = "cf".getBytes();
	public static final byte[] COUNT = "count".getBytes();

 	public void reduce(Text key, Iterable<IntWritable> values, Context context) throws IOException, InterruptedException {
    		int i = 0;
    		for (IntWritable val : values) {
    			i += val.get();
    		Put put = new Put(Bytes.toBytes(key.toString()));
    		put.add(CF, COUNT, Bytes.toBytes(i));

    		context.write(null, put);

7.2.5. HBase MapReduce Summary to File Example

This very similar to the summary example above, with exception that this is using HBase as a MapReduce source but HDFS as the sink. The differences are in the job setup and in the reducer. The mapper remains the same.

Configuration config = HBaseConfiguration.create();
Job job = new Job(config,"ExampleSummaryToFile");
job.setJarByClass(MySummaryFileJob.class);     // class that contains mapper and reducer

Scan scan = new Scan();
scan.setCaching(500);        // 1 is the default in Scan, which will be bad for MapReduce jobs
scan.setCacheBlocks(false);  // don't set to true for MR jobs
// set other scan attrs

	sourceTable,        // input table
	scan,               // Scan instance to control CF and attribute selection
	MyMapper.class,     // mapper class
	Text.class,         // mapper output key
	IntWritable.class,  // mapper output value
job.setReducerClass(MyReducer.class);    // reducer class
job.setNumReduceTasks(1);    // at least one, adjust as required
FileOutputFormat.setOutputPath(job, new Path("/tmp/mr/mySummaryFile"));  // adjust directories as required

boolean b = job.waitForCompletion(true);
if (!b) {
	throw new IOException("error with job!");
As stated above, the previous Mapper can run unchanged with this example. As for the Reducer, it is a "generic" Reducer instead of extending TableMapper and emitting Puts.
 public static class MyReducer extends Reducer<Text, IntWritable, Text, IntWritable>  {

	public void reduce(Text key, Iterable<IntWritable> values, Context context) throws IOException, InterruptedException {
		int i = 0;
		for (IntWritable val : values) {
			i += val.get();
		context.write(key, new IntWritable(i));

7.2.6. HBase MapReduce Summary to HBase Without Reducer

It is also possible to perform summaries without a reducer - if you use HBase as the reducer.

An HBase target table would need to exist for the job summary. The HTable method incrementColumnValue would be used to atomically increment values. From a performance perspective, it might make sense to keep a Map of values with their values to be incremeneted for each map-task, and make one update per key at during the cleanup method of the mapper. However, your milage may vary depending on the number of rows to be processed and unique keys.

In the end, the summary results are in HBase.

7.2.7. HBase MapReduce Summary to RDBMS

Sometimes it is more appropriate to generate summaries to an RDBMS. For these cases, it is possible to generate summaries directly to an RDBMS via a custom reducer. The setup method can connect to an RDBMS (the connection information can be passed via custom parameters in the context) and the cleanup method can close the connection.

It is critical to understand that number of reducers for the job affects the summarization implementation, and you'll have to design this into your reducer. Specifically, whether it is designed to run as a singleton (one reducer) or multiple reducers. Neither is right or wrong, it depends on your use-case. Recognize that the more reducers that are assigned to the job, the more simultaneous connections to the RDBMS will be created - this will scale, but only to a point.

 public static class MyRdbmsReducer extends Reducer<Text, IntWritable, Text, IntWritable>  {

	private Connection c = null;

	public void setup(Context context) {
  		// create DB connection...

	public void reduce(Text key, Iterable<IntWritable> values, Context context) throws IOException, InterruptedException {
		// do summarization
		// in this example the keys are Text, but this is just an example

	public void cleanup(Context context) {
  		// close db connection


In the end, the summary results are written to your RDBMS table/s.

7.3. Accessing Other HBase Tables in a MapReduce Job

Although the framework currently allows one HBase table as input to a MapReduce job, other HBase tables can be accessed as lookup tables, etc., in a MapReduce job via creating an HTable instance in the setup method of the Mapper.

public class MyMapper extends TableMapper<Text, LongWritable> {
  private HTable myOtherTable;

  public void setup(Context context) {
    myOtherTable = new HTable("myOtherTable");

  public void map(ImmutableBytesWritable row, Result value, Context context) throws IOException, InterruptedException {
	// process Result...
	// use 'myOtherTable' for lookups


7.4. Speculative Execution

It is generally advisable to turn off speculative execution for MapReduce jobs that use HBase as a source. This can either be done on a per-Job basis through properties, on on the entire cluster. Especially for longer running jobs, speculative execution will create duplicate map-tasks which will double-write your data to HBase; this is probably not what you want.

See Section, “Speculative Execution” for more information.

Chapter 8. Secure Apache HBase

8.1. Secure Client Access to Apache HBase

Newer releases of Apache HBase (>= 0.92) support optional SASL authentication of clients[22].

This describes how to set up Apache HBase and clients for connection to secure HBase resources.

8.1.1. Prerequisites

You need to have a working Kerberos KDC.

A HBase configured for secure client access is expected to be running on top of a secured HDFS cluster. HBase must be able to authenticate to HDFS services. HBase needs Kerberos credentials to interact with the Kerberos-enabled HDFS daemons. Authenticating a service should be done using a keytab file. The procedure for creating keytabs for HBase service is the same as for creating keytabs for Hadoop. Those steps are omitted here. Copy the resulting keytab files to wherever HBase Master and RegionServer processes are deployed and make them readable only to the user account under which the HBase daemons will run.

A Kerberos principal has three parts, with the form username/ We recommend using hbase as the username portion.

The following is an example of the configuration properties for Kerberos operation that must be added to the hbase-site.xml file on every server machine in the cluster. Required for even the most basic interactions with a secure Hadoop configuration, independent of HBase security.


Each HBase client user should also be given a Kerberos principal. This principal should have a password assigned to it (as opposed to a keytab file). The client principal's maxrenewlife should be set so that it can be renewed enough times for the HBase client process to complete. For example, if a user runs a long-running HBase client process that takes at most 3 days, we might create this user's principal within kadmin with: addprinc -maxrenewlife 3days

Long running daemons with indefinite lifetimes that require client access to HBase can instead be configured to log in from a keytab. For each host running such daemons, create a keytab with kadmin or kadmin.local. The procedure for creating keytabs for HBase service is the same as for creating keytabs for Hadoop. Those steps are omitted here. Copy the resulting keytab files to where the client daemon will execute and make them readable only to the user account under which the daemon will run.

8.1.2. Server-side Configuration for Secure Operation

Add the following to the hbase-site.xml file on every server machine in the cluster:


A full shutdown and restart of HBase service is required when deploying these configuration changes.

8.1.3. Client-side Configuration for Secure Operation

Add the following to the hbase-site.xml file on every client:


The client environment must be logged in to Kerberos from KDC or keytab via the kinit command before communication with the HBase cluster will be possible.

Be advised that if the in the client- and server-side site files do not match, the client will not be able to communicate with the cluster.

Once HBase is configured for secure RPC it is possible to optionally configure encrypted communication. To do so, add the following to the hbase-site.xml file on every client:


This configuration property can also be set on a per connection basis. Set it in the Configuration supplied to HTable:

      Configuration conf = HBaseConfiguration.create();
      conf.set("", "privacy");
      HTable table = new HTable(conf, tablename);

Expect a ~10% performance penalty for encrypted communication.

8.1.4. Client-side Configuration for Secure Operation - Thrift Gateway

Add the following to the hbase-site.xml file for every Thrift gateway:

      <!-- TODO: This may need to be  HTTP/_HOST@<REALM> and _HOST may not work.
       You may have  to put the concrete full hostname.

Substitute the appropriate credential and keytab for $USER and $KEYTAB respectively.

In order to use the Thrift API principal to interact with HBase, it is also necessary to add the hbase.thrift.kerberos.principal to the _acl_ table. For example, to give the Thrift API principal, thrift_server, administrative access, a command such as this one will suffice:

    grant 'thrift_server', 'RWCA'

For more information about ACLs, please see the Access Control section

The Thrift gateway will authenticate with HBase using the supplied credential. No authentication will be performed by the Thrift gateway itself. All client access via the Thrift gateway will use the Thrift gateway's credential and have its privilege.

8.1.5. Client-side Configuration for Secure Operation - REST Gateway

Add the following to the hbase-site.xml file for every REST gateway:


Substitute the appropriate credential and keytab for $USER and $KEYTAB respectively.

The REST gateway will authenticate with HBase using the supplied credential. No authentication will be performed by the REST gateway itself. All client access via the REST gateway will use the REST gateway's credential and have its privilege.

In order to use the REST API principal to interact with HBase, it is also necessary to add the to the _acl_ table. For example, to give the REST API principal, rest_server, administrative access, a command such as this one will suffice:

    grant 'rest_server', 'RWCA'

For more information about ACLs, please see the Access Control section

It should be possible for clients to authenticate with the HBase cluster through the REST gateway in a pass-through manner via SPEGNO HTTP authentication. This is future work.

8.2. Access Control

Newer releases of Apache HBase (>= 0.92) support optional access control list (ACL-) based protection of resources on a column family and/or table basis.

This describes how to set up Secure HBase for access control, with an example of granting and revoking user permission on table resources provided.

8.2.1. Prerequisites

You must configure HBase for secure operation. Refer to the section "Secure Client Access to HBase" and complete all of the steps described there.

You must also configure ZooKeeper for secure operation. Changes to ACLs are synchronized throughout the cluster using ZooKeeper. Secure authentication to ZooKeeper must be enabled or otherwise it will be possible to subvert HBase access control via direct client access to ZooKeeper. Refer to the section on secure ZooKeeper configuration and complete all of the steps described there.

8.2.2. Overview

With Secure RPC and Access Control enabled, client access to HBase is authenticated and user data is private unless access has been explicitly granted. Access to data can be granted at a table or per column family basis.

However, the following items have been left out of the initial implementation for simplicity:

  1. Row-level or per value (cell): This would require broader changes for storing the ACLs inline with rows. It is a future goal.

  2. Push down of file ownership to HDFS: HBase is not designed for the case where files may have different permissions than the HBase system principal. Pushing file ownership down into HDFS would necessitate changes to core code. Also, while HDFS file ownership would make applying quotas easy, and possibly make bulk imports more straightforward, it is not clear that it would offer a more secure setup.

  3. HBase managed "roles" as collections of permissions: We will not model "roles" internally in HBase to begin with. We instead allow group names to be granted permissions, which allows external modeling of roles via group membership. Groups are created and manipulated externally to HBase, via the Hadoop group mapping service.

Access control mechanisms are mature and fairly standardized in the relational database world. The HBase implementation approximates current convention, but HBase has a simpler feature set than relational databases, especially in terms of client operations. We don't distinguish between an insert (new record) and update (of existing record), for example, as both collapse down into a Put. Accordingly, the important operations condense to four permissions: READ, WRITE, CREATE, and ADMIN.

Table 8.1. Operation To Permission Mapping

 Bulk Load
 Major Compact

Permissions can be granted in any of the following scopes, though CREATE and ADMIN permissions are effective only at table scope.

  • Table

    • Read: User can read from any column family in table

    • Write: User can write to any column family in table

    • Create: User can alter table attributes; add, alter, or drop column families; and drop the table.

    • Admin: User can alter table attributes; add, alter, or drop column families; and enable, disable, or drop the table. User can also trigger region (re)assignments or relocation.

  • Column Family

    • Read: User can read from the column family

    • Write: User can write to the column family

There is also an implicit global scope for the superuser.

The superuser is a principal, specified in the HBase site configuration file, that has equivalent access to HBase as the 'root' user would on a UNIX derived system. Normally this is the principal that the HBase processes themselves authenticate as. Although future versions of HBase Access Control may support multiple superusers, the superuser privilege will always include the principal used to run the HMaster process. Only the superuser is allowed to create tables, switch the balancer on or off, or take other actions with global consequence. Furthermore, the superuser has an implicit grant of all permissions to all resources.

Tables have a new metadata attribute: OWNER, the user principal who owns the table. By default this will be set to the user principal who creates the table, though it may be changed at table creation time or during an alter operation by setting or changing the OWNER table attribute. Only a single user principal can own a table at a given time. A table owner will have all permissions over a given table.

8.2.3. Server-side Configuration for Access Control

Enable the AccessController coprocessor in the cluster configuration and restart HBase. The restart can be a rolling one. Complete the restart of all Master and RegionServer processes before setting up ACLs.

To enable the AccessController, modify the hbase-site.xml file on every server machine in the cluster to look like:


8.2.4. Shell Enhancements for Access Control

The HBase shell has been extended to provide simple commands for editing and updating user permissions. The following commands have been added for access control list management:


    grant <user> <permissions> [ <table> [ <column family> [ <column qualifier> ] ] ]

<permissions> is zero or more letters from the set "RWCA": READ('R'), WRITE('W'), CREATE('C'), ADMIN('A').

Note: Grants and revocations of individual permissions on a resource are both accomplished using the grant command. A separate revoke command is also provided by the shell, but this is for fast revocation of all of a user's access rights to a given resource only.


    revoke <user> [ <table> [ <column family> [ <column qualifier> ] ] ]


The alter command has been extended to allow ownership assignment:

      alter 'tablename', {OWNER => 'username'}

User Permission

The user_permission command shows all access permissions for the current user for a given table:

      user_permission <table>

8.3. Secure Bulk Load

Bulk loading in secure mode is a bit more involved than normal setup, since the client has to transfer the ownership of the files generated from the mapreduce job to HBase. Secure bulk loading is implemented by a coprocessor, named SecureBulkLoadEndpoint. SecureBulkLoadEndpoint uses a staging directory "hbase.bulkload.staging.dir", which defaults to /tmp/hbase-staging/. The algorithm is as follows.

  • Create an hbase owned staging directory which is world traversable (-rwx--x--x, 711) /tmp/hbase-staging.
  • A user writes out data to his secure output directory: /user/foo/data
  • A call is made to hbase to create a secret staging directory which is globally readable/writable (-rwxrwxrwx, 777): /tmp/hbase-staging/averylongandrandomdirectoryname
  • The user makes the data world readable and writable, then moves it into the random staging directory, then calls bulkLoadHFiles()

Like delegation tokens the strength of the security lies in the length and randomness of the secret directory.

You have to enable the secure bulk load to work properly. You can modify the hbase-site.xml file on every server machine in the cluster and add the SecureBulkLoadEndpoint class to the list of regionserver coprocessors:


Chapter 9. Architecture

9.1. Overview

9.1.1. NoSQL?

HBase is a type of "NoSQL" database. "NoSQL" is a general term meaning that the database isn't an RDBMS which supports SQL as its primary access language, but there are many types of NoSQL databases: BerkeleyDB is an example of a local NoSQL database, whereas HBase is very much a distributed database. Technically speaking, HBase is really more a "Data Store" than "Data Base" because it lacks many of the features you find in an RDBMS, such as typed columns, secondary indexes, triggers, and advanced query languages, etc.

However, HBase has many features which supports both linear and modular scaling. HBase clusters expand by adding RegionServers that are hosted on commodity class servers. If a cluster expands from 10 to 20 RegionServers, for example, it doubles both in terms of storage and as well as processing capacity. RDBMS can scale well, but only up to a point - specifically, the size of a single database server - and for the best performance requires specialized hardware and storage devices. HBase features of note are:

  • Strongly consistent reads/writes: HBase is not an "eventually consistent" DataStore. This makes it very suitable for tasks such as high-speed counter aggregation.
  • Automatic sharding: HBase tables are distributed on the cluster via regions, and regions are automatically split and re-distributed as your data grows.
  • Automatic RegionServer failover
  • Hadoop/HDFS Integration: HBase supports HDFS out of the box as its distributed file system.
  • MapReduce: HBase supports massively parallelized processing via MapReduce for using HBase as both source and sink.
  • Java Client API: HBase supports an easy to use Java API for programmatic access.
  • Thrift/REST API: HBase also supports Thrift and REST for non-Java front-ends.
  • Block Cache and Bloom Filters: HBase supports a Block Cache and Bloom Filters for high volume query optimization.
  • Operational Management: HBase provides build-in web-pages for operational insight as well as JMX metrics.

9.1.2. When Should I Use HBase?

HBase isn't suitable for every problem.

First, make sure you have enough data. If you have hundreds of millions or billions of rows, then HBase is a good candidate. If you only have a few thousand/million rows, then using a traditional RDBMS might be a better choice due to the fact that all of your data might wind up on a single node (or two) and the rest of the cluster may be sitting idle.

Second, make sure you can live without all the extra features that an RDBMS provides (e.g., typed columns, secondary indexes, transactions, advanced query languages, etc.) An application built against an RDBMS cannot be "ported" to HBase by simply changing a JDBC driver, for example. Consider moving from an RDBMS to HBase as a complete redesign as opposed to a port.

Third, make sure you have enough hardware. Even HDFS doesn't do well with anything less than 5 DataNodes (due to things such as HDFS block replication which has a default of 3), plus a NameNode.

HBase can run quite well stand-alone on a laptop - but this should be considered a development configuration only.

9.1.3. What Is The Difference Between HBase and Hadoop/HDFS?

HDFS is a distributed file system that is well suited for the storage of large files. It's documentation states that it is not, however, a general purpose file system, and does not provide fast individual record lookups in files. HBase, on the other hand, is built on top of HDFS and provides fast record lookups (and updates) for large tables. This can sometimes be a point of conceptual confusion. HBase internally puts your data in indexed "StoreFiles" that exist on HDFS for high-speed lookups. See the Chapter 5, Data Model and the rest of this chapter for more information on how HBase achieves its goals.

9.2. Catalog Tables

The catalog tables -ROOT- and .META. exist as HBase tables. They are filtered out of the HBase shell's list command, but they are in fact tables just like any other.

9.2.1. ROOT

-ROOT- keeps track of where the .META. table is. The -ROOT- table structure is as follows:


  • .META. region key (.META.,,1)


  • info:regioninfo (serialized HRegionInfo instance of .META.)
  • info:server (server:port of the RegionServer holding .META.)
  • info:serverstartcode (start-time of the RegionServer process holding .META.)

9.2.2. META

The .META. table keeps a list of all regions in the system. The .META. table structure is as follows:


  • Region key of the format ([table],[region start key],[region id])


  • info:regioninfo (serialized HRegionInfo instance for this region)
  • info:server (server:port of the RegionServer containing this region)
  • info:serverstartcode (start-time of the RegionServer process containing this region)

When a table is in the process of splitting two other columns will be created, info:splitA and info:splitB which represent the two daughter regions. The values for these columns are also serialized HRegionInfo instances. After the region has been split eventually this row will be deleted.

Notes on HRegionInfo: the empty key is used to denote table start and table end. A region with an empty start key is the first region in a table. If region has both an empty start and an empty end key, it's the only region in the table

In the (hopefully unlikely) event that programmatic processing of catalog metadata is required, see the Writables utility.

9.2.3. Startup Sequencing

The META location is set in ROOT first. Then META is updated with server and startcode values.

For information on region-RegionServer assignment, see Section 9.7.2, “Region-RegionServer Assignment”.

9.3. Client

The HBase client HTable is responsible for finding RegionServers that are serving the particular row range of interest. It does this by querying the .META. and -ROOT- catalog tables (TODO: Explain). After locating the required region(s), the client directly contacts the RegionServer serving that region (i.e., it does not go through the master) and issues the read or write request. This information is cached in the client so that subsequent requests need not go through the lookup process. Should a region be reassigned either by the master load balancer or because a RegionServer has died, the client will requery the catalog tables to determine the new location of the user region.

See Section 9.5.2, “Runtime Impact” for more information about the impact of the Master on HBase Client communication.

Administrative functions are handled through HBaseAdmin

9.3.1. Connections

For connection configuration information, see Section 2.3.4, “Client configuration and dependencies connecting to an HBase cluster”.

HTable instances are not thread-safe. Only one thread use an instance of HTable at any given time. When creating HTable instances, it is advisable to use the same HBaseConfiguration instance. This will ensure sharing of ZooKeeper and socket instances to the RegionServers which is usually what you want. For example, this is preferred:

HBaseConfiguration conf = HBaseConfiguration.create();
HTable table1 = new HTable(conf, "myTable");
HTable table2 = new HTable(conf, "myTable");

as opposed to this:

HBaseConfiguration conf1 = HBaseConfiguration.create();
HTable table1 = new HTable(conf1, "myTable");
HBaseConfiguration conf2 = HBaseConfiguration.create();
HTable table2 = new HTable(conf2, "myTable");

For more information about how connections are handled in the HBase client, see HConnectionManager. Connection Pooling

For applications which require high-end multithreaded access (e.g., web-servers or application servers that may serve many application threads in a single JVM), one solution is HTablePool. But as written currently, it is difficult to control client resource consumption when using HTablePool.

Another solution is to precreate an HConnection using

// Create a connection to the cluster.
HConnection connection = HConnectionManager.createConnection(Configuration);
HTableInterface table = connection.getTable("myTable");
// use table as needed, the table returned is lightweight
// use the connection for other access to the cluster

Constructing HTableInterface implementation is very lightweight and resources are controlled/shared if you go this route.

9.3.2. WriteBuffer and Batch Methods

If Section 12.8.4, “HBase Client: AutoFlush” is turned off on HTable, Puts are sent to RegionServers when the writebuffer is filled. The writebuffer is 2MB by default. Before an HTable instance is discarded, either close() or flushCommits() should be invoked so Puts will not be lost.

Note: htable.delete(Delete); does not go in the writebuffer! This only applies to Puts.

For additional information on write durability, review the ACID semantics page.

For fine-grained control of batching of Puts or Deletes, see the batch methods on HTable.

9.3.3. External Clients

Information on non-Java clients and custom protocols is covered in Chapter 10, Apache HBase External APIs

9.3.4. RowLocks

RowLocks are still in the client API however they are discouraged because if not managed properly these can lock up the RegionServers.

There is an oustanding ticket HBASE-2332 to remove this feature from the client.

9.4. Client Request Filters

Get and Scan instances can be optionally configured with filters which are applied on the RegionServer.

Filters can be confusing because there are many different types, and it is best to approach them by understanding the groups of Filter functionality.

9.4.1. Structural

Structural Filters contain other Filters. FilterList

FilterList represents a list of Filters with a relationship of FilterList.Operator.MUST_PASS_ALL or FilterList.Operator.MUST_PASS_ONE between the Filters. The following example shows an 'or' between two Filters (checking for either 'my value' or 'my other value' on the same attribute).

FilterList list = new FilterList(FilterList.Operator.MUST_PASS_ONE);
SingleColumnValueFilter filter1 = new SingleColumnValueFilter(
	Bytes.toBytes("my value")
SingleColumnValueFilter filter2 = new SingleColumnValueFilter(
	Bytes.toBytes("my other value")

9.4.2. Column Value SingleColumnValueFilter

SingleColumnValueFilter can be used to test column values for equivalence (CompareOp.EQUAL ), inequality (CompareOp.NOT_EQUAL), or ranges (e.g., CompareOp.GREATER). The folowing is example of testing equivalence a column to a String value "my value"...

SingleColumnValueFilter filter = new SingleColumnValueFilter(
	Bytes.toBytes("my value")

9.4.3. Column Value Comparators

There are several Comparator classes in the Filter package that deserve special mention. These Comparators are used in concert with other Filters, such as Section, “SingleColumnValueFilter”. RegexStringComparator

RegexStringComparator supports regular expressions for value comparisons.

RegexStringComparator comp = new RegexStringComparator("my.");   // any value that starts with 'my'
SingleColumnValueFilter filter = new SingleColumnValueFilter(

See the Oracle JavaDoc for supported RegEx patterns in Java. SubstringComparator

SubstringComparator can be used to determine if a given substring exists in a value. The comparison is case-insensitive.

SubstringComparator comp = new SubstringComparator("y val");   // looking for 'my value'
SingleColumnValueFilter filter = new SingleColumnValueFilter(
scan.setFilter(filter); BinaryPrefixComparator

See BinaryPrefixComparator. BinaryComparator

See BinaryComparator.

9.4.4. KeyValue Metadata

As HBase stores data internally as KeyValue pairs, KeyValue Metadata Filters evaluate the existence of keys (i.e., ColumnFamily:Column qualifiers) for a row, as opposed to values the previous section. FamilyFilter

FamilyFilter can be used to filter on the ColumnFamily. It is generally a better idea to select ColumnFamilies in the Scan than to do it with a Filter. QualifierFilter

QualifierFilter can be used to filter based on Column (aka Qualifier) name. ColumnPrefixFilter

ColumnPrefixFilter can be used to filter based on the lead portion of Column (aka Qualifier) names.

A ColumnPrefixFilter seeks ahead to the first column matching the prefix in each row and for each involved column family. It can be used to efficiently get a subset of the columns in very wide rows.

Note: The same column qualifier can be used in different column families. This filter returns all matching columns.

Example: Find all columns in a row and family that start with "abc"

HTableInterface t = ...;
byte[] row = ...;
byte[] family = ...;
byte[] prefix = Bytes.toBytes("abc");
Scan scan = new Scan(row, row); // (optional) limit to one row
scan.addFamily(family); // (optional) limit to one family
Filter f = new ColumnPrefixFilter(prefix);
scan.setBatch(10); // set this if there could be many columns returned
ResultScanner rs = t.getScanner(scan);
for (Result r =; r != null; r = {
  for (KeyValue kv : r.raw()) {
    // each kv represents a column
rs.close(); MultipleColumnPrefixFilter

MultipleColumnPrefixFilter behaves like ColumnPrefixFilter but allows specifying multiple prefixes.

Like ColumnPrefixFilter, MultipleColumnPrefixFilter efficiently seeks ahead to the first column matching the lowest prefix and also seeks past ranges of columns between prefixes. It can be used to efficiently get discontinuous sets of columns from very wide rows.

Example: Find all columns in a row and family that start with "abc" or "xyz"

HTableInterface t = ...;
byte[] row = ...;
byte[] family = ...;
byte[][] prefixes = new byte[][] {Bytes.toBytes("abc"), Bytes.toBytes("xyz")};
Scan scan = new Scan(row, row); // (optional) limit to one row
scan.addFamily(family); // (optional) limit to one family
Filter f = new MultipleColumnPrefixFilter(prefixes);
scan.setBatch(10); // set this if there could be many columns returned
ResultScanner rs = t.getScanner(scan);
for (Result r =; r != null; r = {
  for (KeyValue kv : r.raw()) {
    // each kv represents a column
rs.close(); ColumnRangeFilter

A ColumnRangeFilter allows efficient intra row scanning.

A ColumnRangeFilter can seek ahead to the first matching column for each involved column family. It can be used to efficiently get a 'slice' of the columns of a very wide row. i.e. you have a million columns in a row but you only want to look at columns bbbb-bbdd.

Note: The same column qualifier can be used in different column families. This filter returns all matching columns.

Example: Find all columns in a row and family between "bbbb" (inclusive) and "bbdd" (inclusive)

HTableInterface t = ...;
byte[] row = ...;
byte[] family = ...;
byte[] startColumn = Bytes.toBytes("bbbb");
byte[] endColumn = Bytes.toBytes("bbdd");
Scan scan = new Scan(row, row); // (optional) limit to one row
scan.addFamily(family); // (optional) limit to one family
Filter f = new ColumnRangeFilter(startColumn, true, endColumn, true);
scan.setBatch(10); // set this if there could be many columns returned
ResultScanner rs = t.getScanner(scan);
for (Result r =; r != null; r = {
  for (KeyValue kv : r.raw()) {
    // each kv represents a column

Note: Introduced in HBase 0.92

9.4.5. RowKey RowFilter

It is generally a better idea to use the startRow/stopRow methods on Scan for row selection, however RowFilter can also be used.

9.4.6. Utility FirstKeyOnlyFilter

This is primarily used for rowcount jobs. See FirstKeyOnlyFilter.

9.5. Master

HMaster is the implementation of the Master Server. The Master server is responsible for monitoring all RegionServer instances in the cluster, and is the interface for all metadata changes. In a distributed cluster, the Master typically runs on the Section 9.9.1, “NameNode”[23]

9.5.1. Startup Behavior

If run in a multi-Master environment, all Masters compete to run the cluster. If the active Master loses its lease in ZooKeeper (or the Master shuts down), then then the remaining Masters jostle to take over the Master role.

9.5.2. Runtime Impact

A common dist-list question is what happens to an HBase cluster when the Master goes down. Because the HBase client talks directly to the RegionServers, the cluster can still function in a "steady state." Additionally, per Section 9.2, “Catalog Tables” ROOT and META exist as HBase tables (i.e., are not resident in the Master). However, the Master controls critical functions such as RegionServer failover and completing region splits. So while the cluster can still run for a time without the Master, the Master should be restarted as soon as possible.

9.5.3. Interface

The methods exposed by HMasterInterface are primarily metadata-oriented methods:

  • Table (createTable, modifyTable, removeTable, enable, disable)
  • ColumnFamily (addColumn, modifyColumn, removeColumn)
  • Region (move, assign, unassign)

For example, when the HBaseAdmin method disableTable is invoked, it is serviced by the Master server.

9.5.4. Processes

The Master runs several background threads: LoadBalancer

Periodically, and when there are no regions in transition, a load balancer will run and move regions around to balance the cluster's load. See Section, “Balancer” for configuring this property.

See Section 9.7.2, “Region-RegionServer Assignment” for more information on region assignment. CatalogJanitor

Periodically checks and cleans up the .META. table. See Section 9.2.2, “META” for more information on META.

9.6. RegionServer

HRegionServer is the RegionServer implementation. It is responsible for serving and managing regions. In a distributed cluster, a RegionServer runs on a Section 9.9.2, “DataNode”.

9.6.1. Interface

The methods exposed by HRegionRegionInterface contain both data-oriented and region-maintenance methods:

  • Data (get, put, delete, next, etc.)
  • Region (splitRegion, compactRegion, etc.)

For example, when the HBaseAdmin method majorCompact is invoked on a table, the client is actually iterating through all regions for the specified table and requesting a major compaction directly to each region.

9.6.2. Processes

The RegionServer runs a variety of background threads: CompactSplitThread

Checks for splits and handle minor compactions. MajorCompactionChecker

Checks for major compactions. MemStoreFlusher

Periodically flushes in-memory writes in the MemStore to StoreFiles. LogRoller

Periodically checks the RegionServer's HLog.

9.6.3. Coprocessors

Coprocessors were added in 0.92. There is a thorough Blog Overview of CoProcessors posted. Documentation will eventually move to this reference guide, but the blog is the most current information available at this time.

9.6.4. Block Cache Design

The Block Cache is an LRU cache that contains three levels of block priority to allow for scan-resistance and in-memory ColumnFamilies:

  • Single access priority: The first time a block is loaded from HDFS it normally has this priority and it will be part of the first group to be considered during evictions. The advantage is that scanned blocks are more likely to get evicted than blocks that are getting more usage.
  • Mutli access priority: If a block in the previous priority group is accessed again, it upgrades to this priority. It is thus part of the second group considered during evictions.
  • In-memory access priority: If the block's family was configured to be "in-memory", it will be part of this priority disregarding the number of times it was accessed. Catalog tables are configured like this. This group is the last one considered during evictions.

For more information, see the LruBlockCache source Usage

Block caching is enabled by default for all the user tables which means that any read operation will load the LRU cache. This might be good for a large number of use cases, but further tunings are usually required in order to achieve better performance. An important concept is the working set size, or WSS, which is: "the amount of memory needed to compute the answer to a problem". For a website, this would be the data that's needed to answer the queries over a short amount of time.

The way to calculate how much memory is available in HBase for caching is:

            number of region servers * heap size * hfile.block.cache.size * 0.85

The default value for the block cache is 0.25 which represents 25% of the available heap. The last value (85%) is the default acceptable loading factor in the LRU cache after which eviction is started. The reason it is included in this equation is that it would be unrealistic to say that it is possible to use 100% of the available memory since this would make the process blocking from the point where it loads new blocks. Here are some examples:

  • One region server with the default heap size (1GB) and the default block cache size will have 217MB of block cache available.
  • 20 region servers with the heap size set to 8GB and a default block cache size will have 34GB of block cache.
  • 100 region servers with the heap size set to 24GB and a block cache size of 0.5 will have about 1TB of block cache.

Your data isn't the only resident of the block cache, here are others that you may have to take into account:

  • Catalog tables: The -ROOT- and .META. tables are forced into the block cache and have the in-memory priority which means that they are harder to evict. The former never uses more than a few hundreds of bytes while the latter can occupy a few MBs (depending on the number of regions).
  • HFiles indexes: HFile is the file format that HBase uses to store data in HDFS and it contains a multi-layered index in order seek to the data without having to read the whole file. The size of those indexes is a factor of the block size (64KB by default), the size of your keys and the amount of data you are storing. For big data sets it's not unusual to see numbers around 1GB per region server, although not all of it will be in cache because the LRU will evict indexes that aren't used.
  • Keys: Taking into account only the values that are being stored is missing half the picture since every value is stored along with its keys (row key, family, qualifier, and timestamp). See Section 6.3.2, “Try to minimize row and column sizes”.
  • Bloom filters: Just like the HFile indexes, those data structures (when enabled) are stored in the LRU.

Currently the recommended way to measure HFile indexes and bloom filters sizes is to look at the region server web UI and checkout the relevant metrics. For keys, sampling can be done by using the HFile command line tool and look for the average key size metric.

It's generally bad to use block caching when the WSS doesn't fit in memory. This is the case when you have for example 40GB available across all your region servers' block caches but you need to process 1TB of data. One of the reasons is that the churn generated by the evictions will trigger more garbage collections unnecessarily. Here are two use cases:

  • Fully random reading pattern: This is a case where you almost never access the same row twice within a short amount of time such that the chance of hitting a cached block is close to 0. Setting block caching on such a table is a waste of memory and CPU cycles, more so that it will generate more garbage to pick up by the JVM. For more information on monitoring GC, see Section 13.2.3, “JVM Garbage Collection Logs”.
  • Mapping a table: In a typical MapReduce job that takes a table in input, every row will be read only once so there's no need to put them into the block cache. The Scan object has the option of turning this off via the setCaching method (set it to false). You can still keep block caching turned on on this table if you need fast random read access. An example would be counting the number of rows in a table that serves live traffic, caching every block of that table would create massive churn and would surely evict data that's currently in use.

9.6.5. Write Ahead Log (WAL) Purpose

Each RegionServer adds updates (Puts, Deletes) to its write-ahead log (WAL) first, and then to the Section, “MemStore” for the affected Section 9.7.6, “Store”. This ensures that HBase has durable writes. Without WAL, there is the possibility of data loss in the case of a RegionServer failure before each MemStore is flushed and new StoreFiles are written. HLog is the HBase WAL implementation, and there is one HLog instance per RegionServer.

The WAL is in HDFS in /hbase/.logs/ with subdirectories per region.

For more general information about the concept of write ahead logs, see the Wikipedia Write-Ahead Log article. WAL Flushing

TODO (describe). WAL Splitting How edits are recovered from a crashed RegionServer

When a RegionServer crashes, it will lose its ephemeral lease in ZooKeeper...TODO hbase.hlog.split.skip.errors

When set to true, any error encountered splitting will be logged, the problematic WAL will be moved into the .corrupt directory under the hbase rootdir, and processing will continue. If set to false, the default, the exception will be propagated and the split logged as failed.[24] How EOFExceptions are treated when splitting a crashed RegionServers' WALs

If we get an EOF while splitting logs, we proceed with the split even when hbase.hlog.split.skip.errors == false. An EOF while reading the last log in the set of files to split is near-guaranteed since the RegionServer likely crashed mid-write of a record. But we'll continue even if we got an EOF reading other than the last file in the set.[25]

9.7. Regions

Regions are the basic element of availability and distribution for tables, and are comprised of a Store per Column Family. The heirarchy of objects is as follows:

Table       (HBase table)
    Region       (Regions for the table)
         Store          (Store per ColumnFamily for each Region for the table)
              MemStore           (MemStore for each Store for each Region for the table)
              StoreFile          (StoreFiles for each Store for each Region for the table)
                    Block             (Blocks within a StoreFile within a Store for each Region for the table)

For a description of what HBase files look like when written to HDFS, see Section 13.7.2, “Browsing HDFS for HBase Objects”.

In general, HBase is designed to run with a small (20-200) number of relatively large (5-20Gb) regions per server. The considerations for this are as follows: Why cannot I have too many regions?

Typically you want to keep your region count low on HBase for numerous reasons. Usually right around 100 regions per RegionServer has yielded the best results. Here are some of the reasons below for keeping region count low:

  1. MSLAB requires 2mb per memstore (that's 2mb per family per region). 1000 regions that have 2 families each is 3.9GB of heap used, and it's not even storing data yet. NB: the 2MB value is configurable.

  2. If you fill all the regions at somewhat the same rate, the global memory usage makes it that it forces tiny flushes when you have too many regions which in turn generates compactions. Rewriting the same data tens of times is the last thing you want. An example is filling 1000 regions (with one family) equally and let's consider a lower bound for global memstore usage of 5GB (the region server would have a big heap). Once it reaches 5GB it will force flush the biggest region, at that point they should almost all have about 5MB of data so it would flush that amount. 5MB inserted later, it would flush another region that will now have a bit over 5MB of data, and so on. This is currently the main limiting factor for the number of regions; see Section, “Number of regions per RS - upper bound” for detailed formula.

  3. The master as is is allergic to tons of regions, and will take a lot of time assigning them and moving them around in batches. The reason is that it's heavy on ZK usage, and it's not very async at the moment (could really be improved -- and has been imporoved a bunch in 0.96 hbase).

  4. In older versions of HBase (pre-v2 hfile, 0.90 and previous), tons of regions on a few RS can cause the store file index to rise, increasing heap usage and potentially creating memory pressure or OOME on the RSs

Another issue is the effect of the number of regions on mapreduce jobs; it is typical to have one mapper per HBase region. Thus, hosting only 5 regions per RS may not be enough to get sufficient number of tasks for a mapreduce job, while 1000 regions will generate far too many tasks.

See Section 15.9.2, “Determining region count and size” for configuration guidelines.

9.7.2. Region-RegionServer Assignment

This section describes how Regions are assigned to RegionServers. Startup

When HBase starts regions are assigned as follows (short version):

  1. The Master invokes the AssignmentManager upon startup.
  2. The AssignmentManager looks at the existing region assignments in META.
  3. If the region assignment is still valid (i.e., if the RegionServer is still online) then the assignment is kept.
  4. If the assignment is invalid, then the LoadBalancerFactory is invoked to assign the region. The DefaultLoadBalancer will randomly assign the region to a RegionServer.
  5. META is updated with the RegionServer assignment (if needed) and the RegionServer start codes (start time of the RegionServer process) upon region opening by the RegionServer. Failover

When a RegionServer fails (short version):

  1. The regions immediately become unavailable because the RegionServer is down.
  2. The Master will detect that the RegionServer has failed.
  3. The region assignments will be considered invalid and will be re-assigned just like the startup sequence. Region Load Balancing

Regions can be periodically moved by the Section, “LoadBalancer”.

9.7.3. Region-RegionServer Locality

Over time, Region-RegionServer locality is achieved via HDFS block replication. The HDFS client does the following by default when choosing locations to write replicas:

  1. First replica is written to local node
  2. Second replica is written to another node in same rack
  3. Third replica is written to a node in another rack (if sufficient nodes)

Thus, HBase eventually achieves locality for a region after a flush or a compaction. In a RegionServer failover situation a RegionServer may be assigned regions with non-local StoreFiles (because none of the replicas are local), however as new data is written in the region, or the table is compacted and StoreFiles are re-written, they will become "local" to the RegionServer.

For more information, see HDFS Design on Replica Placement and also Lars George's blog on HBase and HDFS locality.

9.7.4. Region Splits

Splits run unaided on the RegionServer; i.e. the Master does not participate. The RegionServer splits a region, offlines the split region and then adds the daughter regions to META, opens daughters on the parent's hosting RegionServer and then reports the split to the Master. See Section, “Managed Splitting” for how to manually manage splits (and for why you might do this) Custom Split Policies

The default split policy can be overwritten using a custom RegionSplitPolicy (HBase 0.94+). Typically a custom split policy should extend HBase's default split policy: ConstantSizeRegionSplitPolicy.

The policy can set globally through the HBaseConfiguration used or on a per table basis:

HTableDescriptor myHtd = ...;
myHtd.setValue(HTableDescriptor.SPLIT_POLICY, MyCustomSplitPolicy.class.getName());

9.7.5. Online Region Merges

Both Master and Regionserver participate in the event of online region merges. Client sends merge RPC to master, then master moves the regions together to the same regionserver where the more heavily loaded region resided, finally master send merge request to this regionserver and regionserver run the region merges. Similar with process of region splits, region merges run as a local transaction on the regionserver, offlines the regions and then merges two regions on the file system, atomically delete merging regions from META and add merged region to the META, opens merged region on the regionserver and reports the merge to Master at last.

An example of region merges in the hbase shell

          hbase> merge_region 'ENCODED_REGIONNAME', 'ENCODED_REGIONNAME', true

It's an asynchronous operation and call returns immediately without waiting merge completed. Passing 'true' as the optional third parameter will force a merge ('force' merges regardless else merge will fail unless passed adjacent regions. 'force' is for expert use only)

9.7.6. Store

A Store hosts a MemStore and 0 or more StoreFiles (HFiles). A Store corresponds to a column family for a table for a given region. MemStore

The MemStore holds in-memory modifications to the Store. Modifications are KeyValues. When asked to flush, current memstore is moved to snapshot and is cleared. HBase continues to serve edits out of new memstore and backing snapshot until flusher reports in that the flush succeeded. At this point the snapshot is let go. StoreFile (HFile)

StoreFiles are where your data lives. HFile Format

The hfile file format is based on the SSTable file described in the BigTable [2006] paper and on Hadoop's tfile (The unit test suite and the compression harness were taken directly from tfile). Schubert Zhang's blog post on HFile: A Block-Indexed File Format to Store Sorted Key-Value Pairs makes for a thorough introduction to HBase's hfile. Matteo Bertozzi has also put up a helpful description, HBase I/O: HFile.

For more information, see the HFile source code. Also see Appendix E, HFile format version 2 for information about the HFile v2 format that was included in 0.92. HFile Tool

To view a textualized version of hfile content, you can do use the tool. Type the following to see usage:

$ ${HBASE_HOME}/bin/hbase  

For example, to view the content of the file hdfs://, type the following:

 $ ${HBASE_HOME}/bin/hbase -v -f hdfs://  

If you leave off the option -v to see just a summary on the hfile. See usage for other things to do with the HFile tool. StoreFile Directory Structure on HDFS

For more information of what StoreFiles look like on HDFS with respect to the directory structure, see Section 13.7.2, “Browsing HDFS for HBase Objects”. Blocks

StoreFiles are composed of blocks. The blocksize is configured on a per-ColumnFamily basis.

Compression happens at the block level within StoreFiles. For more information on compression, see Appendix C, Compression In HBase.

For more information on blocks, see the HFileBlock source code. KeyValue

The KeyValue class is the heart of data storage in HBase. KeyValue wraps a byte array and takes offsets and lengths into passed array at where to start interpreting the content as KeyValue.

The KeyValue format inside a byte array is:

  • keylength
  • valuelength
  • key
  • value

The Key is further decomposed as:

  • rowlength
  • row (i.e., the rowkey)
  • columnfamilylength
  • columnfamily
  • columnqualifier
  • timestamp
  • keytype (e.g., Put, Delete, DeleteColumn, DeleteFamily)

KeyValue instances are not split across blocks. For example, if there is an 8 MB KeyValue, even if the block-size is 64kb this KeyValue will be read in as a coherent block. For more information, see the KeyValue source code. Example

To emphasize the points above, examine what happens with two Puts for two different columns for the same row:

  • Put #1: rowkey=row1, cf:attr1=value1
  • Put #2: rowkey=row1, cf:attr2=value2

Even though these are for the same row, a KeyValue is created for each column:

Key portion for Put #1:

  • rowlength ------------> 4
  • row -----------------> row1
  • columnfamilylength ---> 2
  • columnfamily --------> cf
  • columnqualifier ------> attr1
  • timestamp -----------> server time of Put
  • keytype -------------> Put

Key portion for Put #2:

  • rowlength ------------> 4
  • row -----------------> row1
  • columnfamilylength ---> 2
  • columnfamily --------> cf
  • columnqualifier ------> attr2
  • timestamp -----------> server time of Put
  • keytype -------------> Put

It is critical to understand that the rowkey, ColumnFamily, and column (aka columnqualifier) are embedded within the KeyValue instance. The longer these identifiers are, the bigger the KeyValue is. Compaction

There are two types of compactions: minor and major. Minor compactions will usually pick up a couple of the smaller adjacent StoreFiles and rewrite them as one. Minors do not drop deletes or expired cells, only major compactions do this. Sometimes a minor compaction will pick up all the StoreFiles in the Store and in this case it actually promotes itself to being a major compaction.

After a major compaction runs there will be a single StoreFile per Store, and this will help performance usually. Caution: major compactions rewrite all of the Stores data and on a loaded system, this may not be tenable; major compactions will usually have to be done manually on large systems. See Section, “Managed Compactions”.

Compactions will not perform region merges. See Section 15.2.2, “Merge” for more information on region merging. Compaction File Selection

To understand the core algorithm for StoreFile selection, there is some ASCII-art in the Store source code that will serve as useful reference. It has been copied below:

/* normal skew:
 *         older ----> newer
 *     _
 *    | |   _
 *    | |  | |   _
 *  --|-|- |-|- |-|---_-------_-------  minCompactSize
 *    | |  | |  | |  | |  _  | |
 *    | |  | |  | |  | | | | | |
 *    | |  | |  | |  | | | | | |

Important knobs:

  • Ratio used in compaction file selection algorithm (default 1.2f).
  • hbase.hstore.compaction.min (.90 hbase.hstore.compactionThreshold) (files) Minimum number of StoreFiles per Store to be selected for a compaction to occur (default 2).
  • hbase.hstore.compaction.max (files) Maximum number of StoreFiles to compact per minor compaction (default 10).
  • hbase.hstore.compaction.min.size (bytes) Any StoreFile smaller than this setting with automatically be a candidate for compaction. Defaults to hbase.hregion.memstore.flush.size (128 mb).
  • hbase.hstore.compaction.max.size (.92) (bytes) Any StoreFile larger than this setting with automatically be excluded from compaction (default Long.MAX_VALUE).

The minor compaction StoreFile selection logic is size based, and selects a file for compaction when the file <= sum(smaller_files) * hbase.hstore.compaction.ratio. Minor Compaction File Selection - Example #1 (Basic Example)

This example mirrors an example from the unit test TestCompactSelection.

  • = 1.0f
  • hbase.hstore.compaction.min = 3 (files)
  • hbase.hstore.compaction.max = 5 (files)
  • hbase.hstore.compaction.min.size = 10 (bytes)
  • hbase.hstore.compaction.max.size = 1000 (bytes)

The following StoreFiles exist: 100, 50, 23, 12, and 12 bytes apiece (oldest to newest). With the above parameters, the files that would be selected for minor compaction are 23, 12, and 12.


  • 100 --> No, because sum(50, 23, 12, 12) * 1.0 = 97.
  • 50 --> No, because sum(23, 12, 12) * 1.0 = 47.
  • 23 --> Yes, because sum(12, 12) * 1.0 = 24.
  • 12 --> Yes, because the previous file has been included, and because this does not exceed the the max-file limit of 5
  • 12 --> Yes, because the previous file had been included, and because this does not exceed the the max-file limit of 5. Minor Compaction File Selection - Example #2 (Not Enough Files To Compact)

This example mirrors an example from the unit test TestCompactSelection.

  • = 1.0f
  • hbase.hstore.compaction.min = 3 (files)
  • hbase.hstore.compaction.max = 5 (files)
  • hbase.hstore.compaction.min.size = 10 (bytes)
  • hbase.hstore.compaction.max.size = 1000 (bytes)

The following StoreFiles exist: 100, 25, 12, and 12 bytes apiece (oldest to newest). With the above parameters, no compaction will be started.


  • 100 --> No, because sum(25, 12, 12) * 1.0 = 47
  • 25 --> No, because sum(12, 12) * 1.0 = 24
  • 12 --> No. Candidate because sum(12) * 1.0 = 12, there are only 2 files to compact and that is less than the threshold of 3
  • 12 --> No. Candidate because the previous StoreFile was, but there are not enough files to compact Minor Compaction File Selection - Example #3 (Limiting Files To Compact)

This example mirrors an example from the unit test TestCompactSelection.

  • = 1.0f
  • hbase.hstore.compaction.min = 3 (files)
  • hbase.hstore.compaction.max = 5 (files)
  • hbase.hstore.compaction.min.size = 10 (bytes)
  • hbase.hstore.compaction.max.size = 1000 (bytes)

The following StoreFiles exist: 7, 6, 5, 4, 3, 2, and 1 bytes apiece (oldest to newest). With the above parameters, the files that would be selected for minor compaction are 7, 6, 5, 4, 3.


  • 7 --> Yes, because sum(6, 5, 4, 3, 2, 1) * 1.0 = 21. Also, 7 is less than the min-size
  • 6 --> Yes, because sum(5, 4, 3, 2, 1) * 1.0 = 15. Also, 6 is less than the min-size.
  • 5 --> Yes, because sum(4, 3, 2, 1) * 1.0 = 10. Also, 5 is less than the min-size.
  • 4 --> Yes, because sum(3, 2, 1) * 1.0 = 6. Also, 4 is less than the min-size.
  • 3 --> Yes, because sum(2, 1) * 1.0 = 3. Also, 3 is less than the min-size.
  • 2 --> No. Candidate because previous file was selected and 2 is less than the min-size, but the max-number of files to compact has been reached.
  • 1 --> No. Candidate because previous file was selected and 1 is less than the min-size, but max-number of files to compact has been reached. Impact of Key Configuration Options A large ratio (e.g., 10) will produce a single giant file. Conversely, a value of .25 will produce behavior similar to the BigTable compaction algorithm - resulting in 4 StoreFiles.

hbase.hstore.compaction.min.size. Because this limit represents the "automatic include" limit for all StoreFiles smaller than this value, this value may need to be adjusted downwards in write-heavy environments where many 1 or 2 mb StoreFiles are being flushed, because every file will be targeted for compaction and the resulting files may still be under the min-size and require further compaction, etc.

9.8. Bulk Loading

9.8.1. Overview

HBase includes several methods of loading data into tables. The most straightforward method is to either use the TableOutputFormat class from a MapReduce job, or use the normal client APIs; however, these are not always the most efficient methods.

The bulk load feature uses a MapReduce job to output table data in HBase's internal data format, and then directly loads the generated StoreFiles into a running cluster. Using bulk load will use less CPU and network resources than simply using the HBase API.

9.8.2. Bulk Load Architecture

The HBase bulk load process consists of two main steps. Preparing data via a MapReduce job

The first step of a bulk load is to generate HBase data files (StoreFiles) from a MapReduce job using HFileOutputFormat. This output format writes out data in HBase's internal storage format so that they can be later loaded very efficiently into the cluster.

In order to function efficiently, HFileOutputFormat must be configured such that each output HFile fits within a single region. In order to do this, jobs whose output will be bulk loaded into HBase use Hadoop's TotalOrderPartitioner class to partition the map output into disjoint ranges of the key space, corresponding to the key ranges of the regions in the table.

HFileOutputFormat includes a convenience function, configureIncrementalLoad(), which automatically sets up a TotalOrderPartitioner based on the current region boundaries of a table. Completing the data load

After the data has been prepared using HFileOutputFormat, it is loaded into the cluster using completebulkload. This command line tool iterates through the prepared data files, and for each one determines the region the file belongs to. It then contacts the appropriate Region Server which adopts the HFile, moving it into its storage directory and making the data available to clients.

If the region boundaries have changed during the course of bulk load preparation, or between the preparation and completion steps, the completebulkloads utility will automatically split the data files into pieces corresponding to the new boundaries. This process is not optimally efficient, so users should take care to minimize the delay between preparing a bulk load and importing it into the cluster, especially if other clients are simultaneously loading data through other means.

9.8.3. Importing the prepared data using the completebulkload tool

After a data import has been prepared, either by using the importtsv tool with the "importtsv.bulk.output" option or by some other MapReduce job using the HFileOutputFormat, the completebulkload tool is used to import the data into the running cluster.

The completebulkload tool simply takes the output path where importtsv or your MapReduce job put its results, and the table name to import into. For example:

$ hadoop jar hbase-VERSION.jar completebulkload [-c /path/to/hbase/config/hbase-site.xml] /user/todd/myoutput mytable

The -c config-file option can be used to specify a file containing the appropriate hbase parameters (e.g., hbase-site.xml) if not supplied already on the CLASSPATH (In addition, the CLASSPATH must contain the directory that has the zookeeper configuration file if zookeeper is NOT managed by HBase).

Note: If the target table does not already exist in HBase, this tool will create the table automatically.

This tool will run quickly, after which point the new data will be visible in the cluster.

9.8.4. See Also

For more information about the referenced utilities, see Section 15.1.10, “ImportTsv” and Section 15.1.11, “CompleteBulkLoad”.

9.8.5. Advanced Usage

Although the importtsv tool is useful in many cases, advanced users may want to generate data programatically, or import data from other formats. To get started doing so, dig into and check the JavaDoc for HFileOutputFormat.

The import step of the bulk load can also be done programatically. See the LoadIncrementalHFiles class for more information.

9.9. HDFS

As HBase runs on HDFS (and each StoreFile is written as a file on HDFS), it is important to have an understanding of the HDFS Architecture especially in terms of how it stores files, handles failovers, and replicates blocks.

See the Hadoop documentation on HDFS Architecture for more information.

9.9.1. NameNode

The NameNode is responsible for maintaining the filesystem metadata. See the above HDFS Architecture link for more information.

9.9.2. DataNode

The DataNodes are responsible for storing HDFS blocks. See the above HDFS Architecture link for more information.

[23] J Mohamed Zahoor goes into some more detail on the Master Architecture in this blog posting, HBase HMaster Architecture .

[24] See HBASE-2958 When hbase.hlog.split.skip.errors is set to false, we fail the split but thats it. We need to do more than just fail split if this flag is set.

Chapter 10. Apache HBase External APIs

This chapter will cover access to Apache HBase either through non-Java languages, or through custom protocols.

10.1. Non-Java Languages Talking to the JVM

Currently the documentation on this topic in the Apache HBase Wiki. See also the Thrift API Javadoc.

10.2. REST

Currently most of the documentation on REST exists in the Apache HBase Wiki on REST (The REST gateway used to be called 'Stargate'). There are also a nice set of blogs on How-to: Use the Apache HBase REST Interface by Jesse Anderson.

10.3. Thrift

Currently most of the documentation on Thrift exists in the Apache HBase Wiki on Thrift.

10.3.1. Filter Language Use Case


this feature was introduced in Apache HBase 0.92

This allows the user to perform server-side filtering when accessing HBase over Thrift (or in the shell -- see the 'scan' help in the shell). The user specifies a filter via a string. The string is parsed on the server to construct the filter General Filter String Syntax

A simple filter expression is expressed as: “FilterName (argument, argument, ... , argument)”

You must specify the name of the filter followed by the argument list in parenthesis. Commas separate the individual arguments

If the argument represents a string, it should be enclosed in single quotes.

If it represents a boolean, an integer or a comparison operator like <, >, != etc. it should not be enclosed in quotes

The filter name must be one word. All ASCII characters are allowed except for whitespace, single quotes and parenthesis.

The filter’s arguments can contain any ASCII character. If single quotes are present in the argument, they must be escaped by a preceding single quote Compound Filters and Operators

Currently, two binary operators – AND/OR and two unary operators – WHILE/SKIP are supported.

Note: the operators are all in uppercase

AND – as the name suggests, if this operator is used, the key-value must pass both the filters

OR – as the name suggests, if this operator is used, the key-value must pass at least one of the filters

SKIP – For a particular row, if any of the key-values don’t pass the filter condition, the entire row is skipped

WHILE - For a particular row, it continues to emit key-values until a key-value is reached that fails the filter condition

Compound Filters: Using these operators, a hierarchy of filters can be created. For example: “(Filter1 AND Filter2) OR (Filter3 AND Filter4)” Order of Evaluation

Parenthesis have the highest precedence. The SKIP and WHILE operators are next and have the same precedence.The AND operator has the next highest precedence followed by the OR operator.

For example:

A filter string of the form:“Filter1 AND Filter2 OR Filter3” will be evaluated as:“(Filter1 AND Filter2) OR Filter3”

A filter string of the form:“Filter1 AND SKIP Filter2 OR Filter3” will be evaluated as:“(Filter1 AND (SKIP Filter2)) OR Filter3” Compare Operator

A compare operator can be any of the following:

  1. LESS (<)

  2. LESS_OR_EQUAL (<=)

  3. EQUAL (=)

  4. NOT_EQUAL (!=)


  6. GREATER (>)

  7. NO_OP (no operation)

The client should use the symbols (<, <=, =, !=, >, >=) to express compare operators. Comparator

A comparator can be any of the following:

  1. BinaryComparator - This lexicographically compares against the specified byte array using Bytes.compareTo(byte[], byte[])

  2. BinaryPrefixComparator - This lexicographically compares against a specified byte array. It only compares up to the length of this byte array.

  3. RegexStringComparator - This compares against the specified byte array using the given regular expression. Only EQUAL and NOT_EQUAL comparisons are valid with this comparator

  4. SubStringComparator - This tests if the given substring appears in a specified byte array. The comparison is case insensitive. Only EQUAL and NOT_EQUAL comparisons are valid with this comparator

The general syntax of a comparator is: ComparatorType:ComparatorValue

The ComparatorType for the various comparators is as follows:

  1. BinaryComparator - binary

  2. BinaryPrefixComparator - binaryprefix

  3. RegexStringComparator - regexstring

  4. SubStringComparator - substring

The ComparatorValue can be any value.

Example1: >, 'binary:abc' will match everything that is lexicographically greater than "abc"

Example2: =, 'binaryprefix:abc' will match everything whose first 3 characters are lexicographically equal to "abc"

Example3: !=, 'regexstring:ab*yz' will match everything that doesn't begin with "ab" and ends with "yz"

Example4: =, 'substring:abc123' will match everything that begins with the substring "abc123" Example PHP Client Program that uses the Filter Language

<? $_SERVER['PHP_ROOT'] = realpath(dirname(__FILE__).'/..');
   require_once $_SERVER['PHP_ROOT'].'/flib/__flib.php';
   $hbase = new HBase('<server_name_running_thrift_server>', <port on which thrift server is running>);
   $client = $hbase->getClient();
   $result = $client->scannerOpenWithFilterString('table_name', "(PrefixFilter ('row2') AND (QualifierFilter (>=, 'binary:xyz'))) AND (TimestampsFilter ( 123, 456))");
   $to_print = $client->scannerGetList($result,1);
   while ($to_print) {
      $to_print = $client->scannerGetList($result,1);
 Example Filter Strings

  • “PrefixFilter (‘Row’) AND PageFilter (1) AND FirstKeyOnlyFilter ()” will return all key-value pairs that match the following conditions:

    1) The row containing the key-value should have prefix “Row”

    2) The key-value must be located in the first row of the table

    3) The key-value pair must be the first key-value in the row

  • “(RowFilter (=, ‘binary:Row 1’) AND TimeStampsFilter (74689, 89734)) OR ColumnRangeFilter (‘abc’, true, ‘xyz’, false))” will return all key-value pairs that match both the following conditions:

    1) The key-value is in a row having row key “Row 1”

    2) The key-value must have a timestamp of either 74689 or 89734.

    Or it must match the following condition:

    1) The key-value pair must be in a column that is lexicographically >= abc and < xyz 

    • “SKIP ValueFilter (0)” will skip the entire row if any of the values in the row is not 0 Individual Filter Syntax

    1. KeyOnlyFilter

      Description: This filter doesn’t take any arguments. It returns only the key component of each key-value.

      Syntax: KeyOnlyFilter ()

      Example: "KeyOnlyFilter ()"

    2. FirstKeyOnlyFilter

      Description: This filter doesn’t take any arguments. It returns only the first key-value from each row.

      Syntax: FirstKeyOnlyFilter ()

      Example: "FirstKeyOnlyFilter ()"

    3. PrefixFilter

      Description: This filter takes one argument – a prefix of a row key. It returns only those key-values present in a row that starts with the specified row prefix

      Syntax: PrefixFilter (‘<row_prefix>’)

      Example: "PrefixFilter (‘Row’)"

    4. ColumnPrefixFilter

      Description: This filter takes one argument – a column prefix. It returns only those key-values present in a column that starts with the specified column prefix. The column prefix must be of the form: “qualifier”


      Example: "ColumnPrefixFilter(‘Col’)"

    5. MultipleColumnPrefixFilter

      Description: This filter takes a list of column prefixes. It returns key-values that are present in a column that starts with any of the specified column prefixes. Each of the column prefixes must be of the form: “qualifier”

      Syntax:MultipleColumnPrefixFilter(‘<column_prefix>’, ‘<column_prefix>’, …, ‘<column_prefix>’)

      Example: "MultipleColumnPrefixFilter(‘Col1’, ‘Col2’)"

    6. ColumnCountGetFilter

      Description: This filter takes one argument – a limit. It returns the first limit number of columns in the table

      Syntax: ColumnCountGetFilter (‘<limit>’)

      Example: "ColumnCountGetFilter (4)"

    7. PageFilter

      Description: This filter takes one argument – a page size. It returns page size number of rows from the table.

      Syntax: PageFilter (‘<page_size>’)

      Example: "PageFilter (2)"

    8. ColumnPaginationFilter

      Description: This filter takes two arguments – a limit and offset. It returns limit number of columns after offset number of columns. It does this for all the rows

      Syntax: ColumnPaginationFilter(‘<limit>’, ‘<offest>’)

      Example: "ColumnPaginationFilter (3, 5)"

    9. InclusiveStopFilter

      Description: This filter takes one argument – a row key on which to stop scanning. It returns all key-values present in rows up to and including the specified row

      Syntax: InclusiveStopFilter(‘<stop_row_key>’)

      Example: "InclusiveStopFilter ('Row2')"

    10. TimeStampsFilter

      Description: This filter takes a list of timestamps. It returns those key-values whose timestamps matches any of the specified timestamps

      Syntax: TimeStampsFilter (<timestamp>, <timestamp>, ... ,<timestamp>)

      Example: "TimeStampsFilter (5985489, 48895495, 58489845945)"

    11. RowFilter

      Description: This filter takes a compare operator and a comparator. It compares each row key with the comparator using the compare operator and if the comparison returns true, it returns all the key-values in that row

      Syntax: RowFilter (<compareOp>, ‘<row_comparator>’)

      Example: "RowFilter (<=, ‘xyz)"

    12. Family Filter

      Description: This filter takes a compare operator and a comparator. It compares each qualifier name with the comparator using the compare operator and if the comparison returns true, it returns all the key-values in that column

      Syntax: QualifierFilter (<compareOp>, ‘<qualifier_comparator>’)

      Example: "QualifierFilter (=, ‘Column1’)"

    13. QualifierFilter

      Description: This filter takes a compare operator and a comparator. It compares each qualifier name with the comparator using the compare operator and if the comparison returns true, it returns all the key-values in that column

      Syntax: QualifierFilter (<compareOp>,‘<qualifier_comparator>’)

      Example: "QualifierFilter (=,‘Column1’)"

    14. ValueFilter

      Description: This filter takes a compare operator and a comparator. It compares each value with the comparator using the compare operator and if the comparison returns true, it returns that key-value

      Syntax: ValueFilter (<compareOp>,‘<value_comparator>’)

      Example: "ValueFilter (!=, ‘Value’)"

    15. DependentColumnFilter

      Description: This filter takes two arguments – a family and a qualifier. It tries to locate this column in each row and returns all key-values in that row that have the same timestamp. If the row doesn’t contain the specified column – none of the key-values in that row will be returned.

      The filter can also take an optional boolean argument – dropDependentColumn. If set to true, the column we were depending on doesn’t get returned.

      The filter can also take two more additional optional arguments – a compare operator and a value comparator, which are further checks in addition to the family and qualifier. If the dependent column is found, its value should also pass the value check and then only is its timestamp taken into consideration

      Syntax: DependentColumnFilter (‘<family>’, ‘<qualifier>’, <boolean>, <compare operator>, ‘<value comparator’)

      Syntax: DependentColumnFilter (‘<family>’, ‘<qualifier>’, <boolean>)

      Syntax: DependentColumnFilter (‘<family>’, ‘<qualifier>’)

      Example: "DependentColumnFilter (‘conf’, ‘blacklist’, false, >=, ‘zebra’)"

      Example: "DependentColumnFilter (‘conf’, 'blacklist', true)"

      Example: "DependentColumnFilter (‘conf’, 'blacklist')"

    16. SingleColumnValueFilter

      Description: This filter takes a column family, a qualifier, a compare operator and a comparator. If the specified column is not found – all the columns of that row will be emitted. If the column is found and the comparison with the comparator returns true, all the columns of the row will be emitted. If the condition fails, the row will not be emitted.

      This filter also takes two additional optional boolean arguments – filterIfColumnMissing and setLatestVersionOnly

      If the filterIfColumnMissing flag is set to true the columns of the row will not be emitted if the specified column to check is not found in the row. The default value is false.

      If the setLatestVersionOnly flag is set to false, it will test previous versions (timestamps) too. The default value is true.

      These flags are optional and if you must set neither or both

      Syntax: SingleColumnValueFilter(‘<family>’, ‘<qualifier>’, <compare operator>, ‘<comparator>’, <filterIfColumnMissing_boolean>, <latest_version_boolean>)

      Syntax: SingleColumnValueFilter(‘<family>’, ‘<qualifier>, <compare operator>, ‘<comparator>’)

      Example: "SingleColumnValueFilter (‘FamilyA’, ‘Column1’, <=, ‘abc’, true, false)"

      Example: "SingleColumnValueFilter (‘FamilyA’, ‘Column1’, <=, ‘abc’)"

    17. SingleColumnValueExcludeFilter

      Description: This filter takes the same arguments and behaves same as SingleColumnValueFilter – however, if the column is found and the condition passes, all the columns of the row will be emitted except for the tested column value.

      Syntax: SingleColumnValueExcludeFilter('<family>', '<qualifier>', <compare operator>, '<comparator>', <latest_version_boolean>, <filterIfColumnMissing_boolean>)

      Syntax: SingleColumnValueExcludeFilter('<family>', '<qualifier>', <compare operator>, '<comparator>')

      Example: "SingleColumnValueExcludeFilter (‘FamilyA’, ‘Column1’, ‘<=’, ‘abc’, ‘false’, ‘true’)"

      Example: "SingleColumnValueExcludeFilter (‘FamilyA’, ‘Column1’, ‘<=’, ‘abc’)"

    18. ColumnRangeFilter

      Description: This filter is used for selecting only those keys with columns that are between minColumn and maxColumn. It also takes two boolean variables to indicate whether to include the minColumn and maxColumn or not.

      If you don’t want to set the minColumn or the maxColumn – you can pass in an empty argument.

      Syntax: ColumnRangeFilter (‘<minColumn>’, <minColumnInclusive_bool>, ‘<maxColumn>’, <maxColumnInclusive_bool>)

      Example: "ColumnRangeFilter (‘abc’, true, ‘xyz’, false)"

    10.4. C/C++ Apache HBase Client

    FB's Chip Turner wrote a pure C/C++ client. Check it out.

    Chapter 11. Apache HBase Coprocessors

    The idea of HBase coprocessors was inspired by Google's BigTable coprocessors. The Apache HBase Blog on Coprocessor is a very good documentation on that. It has detailed information about the coprocessor framework, terminology, management, and so on.

    Chapter 12. Apache HBase Performance Tuning

    Table of Contents

    12.1. Operating System
    12.1.1. Memory
    12.1.2. 64-bit
    12.1.3. Swapping
    12.2. Network
    12.2.1. Single Switch
    12.2.2. Multiple Switches
    12.2.3. Multiple Racks
    12.2.4. Network Interfaces
    12.3. Java
    12.3.1. The Garbage Collector and Apache HBase
    12.4. HBase Configurations
    12.4.1. Managing Compactions
    12.4.2. hbase.regionserver.handler.count
    12.4.3. hfile.block.cache.size
    12.4.6. hbase.hstore.blockingStoreFiles
    12.4.7. hbase.hregion.memstore.block.multiplier
    12.4.8. hbase.regionserver.checksum.verify
    12.5. ZooKeeper
    12.6. Schema Design
    12.6.1. Number of Column Families
    12.6.2. Key and Attribute Lengths
    12.6.3. Table RegionSize
    12.6.4. Bloom Filters
    12.6.5. ColumnFamily BlockSize
    12.6.6. In-Memory ColumnFamilies
    12.6.7. Compression
    12.7. HBase General Patterns
    12.7.1. Constants
    12.8. Writing to HBase
    12.8.1. Batch Loading
    12.8.2. Table Creation: Pre-Creating Regions
    12.8.3. Table Creation: Deferred Log Flush
    12.8.4. HBase Client: AutoFlush
    12.8.5. HBase Client: Turn off WAL on Puts
    12.8.6. HBase Client: Group Puts by RegionServer
    12.8.7. MapReduce: Skip The Reducer
    12.8.8. Anti-Pattern: One Hot Region
    12.9. Reading from HBase
    12.9.1. Scan Caching
    12.9.2. Scan Attribute Selection
    12.9.3. MapReduce - Input Splits
    12.9.4. Close ResultScanners
    12.9.5. Block Cache
    12.9.6. Optimal Loading of Row Keys
    12.9.7. Concurrency: Monitor Data Spread
    12.9.8. Bloom Filters
    12.10. Deleting from HBase
    12.10.1. Using HBase Tables as Queues
    12.10.2. Delete RPC Behavior
    12.11. HDFS
    12.11.1. Current Issues With Low-Latency Reads
    12.11.2. Leveraging local data
    12.11.3. Performance Comparisons of HBase vs. HDFS
    12.12. Amazon EC2
    12.13. Collocating HBase and MapReduce
    12.14. Case Studies

    12.1. Operating System

    12.1.1. Memory

    RAM, RAM, RAM. Don't starve HBase.

    12.1.2. 64-bit

    Use a 64-bit platform (and 64-bit JVM).

    12.1.3. Swapping

    Watch out for swapping. Set swappiness to 0.

    12.2. Network

    Perhaps the most important factor in avoiding network issues degrading Hadoop and HBbase performance is the switching hardware that is used, decisions made early in the scope of the project can cause major problems when you double or triple the size of your cluster (or more).

    Important items to consider:

    • Switching capacity of the device
    • Number of systems connected
    • Uplink capacity

    12.2.1. Single Switch

    The single most important factor in this configuration is that the switching capacity of the hardware is capable of handling the traffic which can be generated by all systems connected to the switch. Some lower priced commodity hardware can have a slower switching capacity than could be utilized by a full switch.

    12.2.2. Multiple Switches

    Multiple switches are a potential pitfall in the architecture. The most common configuration of lower priced hardware is a simple 1Gbps uplink from one switch to another. This often overlooked pinch point can easily become a bottleneck for cluster communication. Especially with MapReduce jobs that are both reading and writing a lot of data the communication across this uplink could be saturated.

    Mitigation of this issue is fairly simple and can be accomplished in multiple ways:

    • Use appropriate hardware for the scale of the cluster which you're attempting to build.
    • Use larger single switch configurations i.e. single 48 port as opposed to 2x 24 port
    • Configure port trunking for uplinks to utilize multiple interfaces to increase cross switch bandwidth.

    12.2.3. Multiple Racks

    Multiple rack configurations carry the same potential issues as multiple switches, and can suffer performance degradation from two main areas:

    • Poor switch capacity performance
    • Insufficient uplink to another rack

    If the the switches in your rack have appropriate switching capacity to handle all the hosts at full speed, the next most likely issue will be caused by homing more of your cluster across racks. The easiest way to avoid issues when spanning multiple racks is to use port trunking to create a bonded uplink to other racks. The downside of this method however, is in the overhead of ports that could potentially be used. An example of this is, creating an 8Gbps port channel from rack A to rack B, using 8 of your 24 ports to communicate between racks gives you a poor ROI, using too few however can mean you're not getting the most out of your cluster.

    Using 10Gbe links between racks will greatly increase performance, and assuming your switches support a 10Gbe uplink or allow for an expansion card will allow you to save your ports for machines as opposed to uplinks.

    12.2.4. Network Interfaces

    Are all the network interfaces functioning correctly? Are you sure? See the Troubleshooting Case Study in Section 14.3.1, “Case Study #1 (Performance Issue On A Single Node)”.

    12.3. Java

    12.3.1. The Garbage Collector and Apache HBase Long GC pauses

    In his presentation, Avoiding Full GCs with MemStore-Local Allocation Buffers, Todd Lipcon describes two cases of stop-the-world garbage collections common in HBase, especially during loading; CMS failure modes and old generation heap fragmentation brought. To address the first, start the CMS earlier than default by adding -XX:CMSInitiatingOccupancyFraction and setting it down from defaults. Start at 60 or 70 percent (The lower you bring down the threshold, the more GCing is done, the more CPU used). To address the second fragmentation issue, Todd added an experimental facility, , that must be explicitly enabled in Apache HBase 0.90.x (Its defaulted to be on in Apache 0.92.x HBase). See hbase.hregion.memstore.mslab.enabled to true in your Configuration. See the cited slides for background and detail[26]. Be aware that when enabled, each MemStore instance will occupy at least an MSLAB instance of memory. If you have thousands of regions or lots of regions each with many column families, this allocation of MSLAB may be responsible for a good portion of your heap allocation and in an extreme case cause you to OOME. Disable MSLAB in this case, or lower the amount of memory it uses or float less regions per server.

    For more information about GC logs, see Section 13.2.3, “JVM Garbage Collection Logs”.

    12.4. HBase Configurations

    See Section 2.5.2, “Recommended Configurations”.

    12.4.1. Managing Compactions

    For larger systems, managing compactions and splits may be something you want to consider.

    12.4.2. hbase.regionserver.handler.count

    See hbase.regionserver.handler.count.

    12.4.3. hfile.block.cache.size

    See hfile.block.cache.size. A memory setting for the RegionServer process.


    See This memory setting is often adjusted for the RegionServer process depending on needs.


    See This memory setting is often adjusted for the RegionServer process depending on needs.

    12.4.6. hbase.hstore.blockingStoreFiles

    See hbase.hstore.blockingStoreFiles. If there is blocking in the RegionServer logs, increasing this can help.

    12.4.7. hbase.hregion.memstore.block.multiplier

    See hbase.hregion.memstore.block.multiplier. If there is enough RAM, increasing this can help.

    12.4.8. hbase.regionserver.checksum.verify

    Have HBase write the checksum into the datablock and save having to do the checksum seek whenever you read.

    See hbase.regionserver.checksum.verify, hbase.hstore.bytes.per.checksum and hbase.hstore.checksum.algorithm For more information see the release note on HBASE-5074 support checksums in HBase block cache.

    12.5. ZooKeeper

    See Chapter 17, ZooKeeper for information on configuring ZooKeeper, and see the part about having a dedicated disk.

    12.6. Schema Design

    12.6.1. Number of Column Families

    See Section 6.2, “ On the number of column families ”.

    12.6.2. Key and Attribute Lengths

    See Section 6.3.2, “Try to minimize row and column sizes”. See also Section, “However...” for compression caveats.

    12.6.3. Table RegionSize

    The regionsize can be set on a per-table basis via setFileSize on HTableDescriptor in the event where certain tables require different regionsizes than the configured default regionsize.

    See Section 15.9.2, “Determining region count and size” for more information.

    12.6.4. Bloom Filters

    Bloom Filters can be enabled per-ColumnFamily. Use HColumnDescriptor.setBloomFilterType(NONE | ROW | ROWCOL) to enable blooms per Column Family. Default = NONE for no bloom filters. If ROW, the hash of the row will be added to the bloom on each insert. If ROWCOL, the hash of the row + column family + column family qualifier will be added to the bloom on each key insert.

    See HColumnDescriptor and Section 12.9.8, “Bloom Filters” for more information or this answer up in quora, How are bloom filters used in HBase?.

    12.6.5. ColumnFamily BlockSize

    The blocksize can be configured for each ColumnFamily in a table, and this defaults to 64k. Larger cell values require larger blocksizes. There is an inverse relationship between blocksize and the resulting StoreFile indexes (i.e., if the blocksize is doubled then the resulting indexes should be roughly halved).

    See HColumnDescriptor and Section 9.7.6, “Store”for more information.

    12.6.6. In-Memory ColumnFamilies

    ColumnFamilies can optionally be defined as in-memory. Data is still persisted to disk, just like any other ColumnFamily. In-memory blocks have the highest priority in the Section 9.6.4, “Block Cache”, but it is not a guarantee that the entire table will be in memory.

    See HColumnDescriptor for more information.

    12.6.7. Compression

    Production systems should use compression with their ColumnFamily definitions. See Appendix C, Compression In HBase for more information. However...

    Compression deflates data on disk. When it's in-memory (e.g., in the MemStore) or on the wire (e.g., transferring between RegionServer and Client) it's inflated. So while using ColumnFamily compression is a best practice, but it's not going to completely eliminate the impact of over-sized Keys, over-sized ColumnFamily names, or over-sized Column names.

    See Section 6.3.2, “Try to minimize row and column sizes” on for schema design tips, and Section, “KeyValue” for more information on HBase stores data internally.

    12.7. HBase General Patterns

    12.7.1. Constants

    When people get started with HBase they have a tendency to write code that looks like this:

    Get get = new Get(rowkey);
    Result r = htable.get(get);
    byte[] b = r.getValue(Bytes.toBytes("cf"), Bytes.toBytes("attr"));  // returns current version of value

    But especially when inside loops (and MapReduce jobs), converting the columnFamily and column-names to byte-arrays repeatedly is surprisingly expensive. It's better to use constants for the byte-arrays, like this:

    public static final byte[] CF = "cf".getBytes();
    public static final byte[] ATTR = "attr".getBytes();
    Get get = new Get(rowkey);
    Result r = htable.get(get);
    byte[] b = r.getValue(CF, ATTR);  // returns current version of value

    12.8. Writing to HBase

    12.8.1. Batch Loading

    Use the bulk load tool if you can. See Section 9.8, “Bulk Loading”. Otherwise, pay attention to the below.

    12.8.2.  Table Creation: Pre-Creating Regions

    Tables in HBase are initially created with one region by default. For bulk imports, this means that all clients will write to the same region until it is large enough to split and become distributed across the cluster. A useful pattern to speed up the bulk import process is to pre-create empty regions. Be somewhat conservative in this, because too-many regions can actually degrade performance.

    There are two different approaches to pre-creating splits. The first approach is to rely on the default HBaseAdmin strategy (which is implemented in Bytes.split)...

    byte[] startKey = ...;   	// your lowest keuy
    byte[] endKey = ...;   		// your highest key
    int numberOfRegions = ...;	// # of regions to create
    admin.createTable(table, startKey, endKey, numberOfRegions);

    And the other approach is to define the splits yourself...

    byte[][] splits = ...;   // create your own splits
    admin.createTable(table, splits);

    See Section 6.3.6, “Relationship Between RowKeys and Region Splits” for issues related to understanding your keyspace and pre-creating regions.

    12.8.3.  Table Creation: Deferred Log Flush

    The default behavior for Puts using the Write Ahead Log (WAL) is that HLog edits will be written immediately. If deferred log flush is used, WAL edits are kept in memory until the flush period. The benefit is aggregated and asynchronous HLog- writes, but the potential downside is that if the RegionServer goes down the yet-to-be-flushed edits are lost. This is safer, however, than not using WAL at all with Puts.

    Deferred log flush can be configured on tables via HTableDescriptor. The default value of hbase.regionserver.optionallogflushinterval is 1000ms.

    12.8.4. HBase Client: AutoFlush

    When performing a lot of Puts, make sure that setAutoFlush is set to false on your HTable instance. Otherwise, the Puts will be sent one at a time to the RegionServer. Puts added via htable.add(Put) and htable.add( <List> Put) wind up in the same write buffer. If autoFlush = false, these messages are not sent until the write-buffer is filled. To explicitly flush the messages, call flushCommits. Calling close on the HTable instance will invoke flushCommits.

    12.8.5. HBase Client: Turn off WAL on Puts

    A frequently discussed option for increasing throughput on Puts is to call writeToWAL(false). Turning this off means that the RegionServer will not write the Put to the Write Ahead Log, only into the memstore, HOWEVER the consequence is that if there is a RegionServer failure there will be data loss. If writeToWAL(false) is used, do so with extreme caution. You may find in actuality that it makes little difference if your load is well distributed across the cluster.

    In general, it is best to use WAL for Puts, and where loading throughput is a concern to use bulk loading techniques instead.

    12.8.6. HBase Client: Group Puts by RegionServer

    In addition to using the writeBuffer, grouping Puts by RegionServer can reduce the number of client RPC calls per writeBuffer flush. There is a utility HTableUtil currently on TRUNK that does this, but you can either copy that or implement your own verison for those still on 0.90.x or earlier.

    12.8.7. MapReduce: Skip The Reducer

    When writing a lot of data to an HBase table from a MR job (e.g., with TableOutputFormat), and specifically where Puts are being emitted from the Mapper, skip the Reducer step. When a Reducer step is used, all of the output (Puts) from the Mapper will get spooled to disk, then sorted/shuffled to other Reducers that will most likely be off-node. It's far more efficient to just write directly to HBase.

    For summary jobs where HBase is used as a source and a sink, then writes will be coming from the Reducer step (e.g., summarize values then write out result). This is a different processing problem than from the the above case.

    12.8.8. Anti-Pattern: One Hot Region

    If all your data is being written to one region at a time, then re-read the section on processing timeseries data.

    Also, if you are pre-splitting regions and all your data is still winding up in a single region even though your keys aren't monotonically increasing, confirm that your keyspace actually works with the split strategy. There are a variety of reasons that regions may appear "well split" but won't work with your data. As the HBase client communicates directly with the RegionServers, this can be obtained via HTable.getRegionLocation.

    See Section 12.8.2, “ Table Creation: Pre-Creating Regions ”, as well as Section 12.4, “HBase Configurations”

    12.9. Reading from HBase

    12.9.1. Scan Caching

    If HBase is used as an input source for a MapReduce job, for example, make sure that the input Scan instance to the MapReduce job has setCaching set to something greater than the default (which is 1). Using the default value means that the map-task will make call back to the region-server for every record processed. Setting this value to 500, for example, will transfer 500 rows at a time to the client to be processed. There is a cost/benefit to have the cache value be large because it costs more in memory for both client and RegionServer, so bigger isn't always better. Scan Caching in MapReduce Jobs

    Scan settings in MapReduce jobs deserve special attention. Timeouts can result (e.g., UnknownScannerException) in Map tasks if it takes longer to process a batch of records before the client goes back to the RegionServer for the next set of data. This problem can occur because there is non-trivial processing occuring per row. If you process rows quickly, set caching higher. If you process rows more slowly (e.g., lots of transformations per row, writes), then set caching lower.

    Timeouts can also happen in a non-MapReduce use case (i.e., single threaded HBase client doing a Scan), but the processing that is often performed in MapReduce jobs tends to exacerbate this issue.

    12.9.2. Scan Attribute Selection

    Whenever a Scan is used to process large numbers of rows (and especially when used as a MapReduce source), be aware of which attributes are selected. If scan.addFamily is called then all of the attributes in the specified ColumnFamily will be returned to the client. If only a small number of the available attributes are to be processed, then only those attributes should be specified in the input scan because attribute over-selection is a non-trivial performance penalty over large datasets.

    12.9.3. MapReduce - Input Splits

    For MapReduce jobs that use HBase tables as a source, if there a pattern where the "slow" map tasks seem to have the same Input Split (i.e., the RegionServer serving the data), see the Troubleshooting Case Study in Section 14.3.1, “Case Study #1 (Performance Issue On A Single Node)”.

    12.9.4. Close ResultScanners

    This isn't so much about improving performance but rather avoiding performance problems. If you forget to close ResultScanners you can cause problems on the RegionServers. Always have ResultScanner processing enclosed in try/catch blocks...

    Scan scan = new Scan();
    // set attrs...
    ResultScanner rs = htable.getScanner(scan);
    try {
      for (Result r =; r != null; r = {
      // process result...
    } finally {
      rs.close();  // always close the ResultScanner!

    12.9.5. Block Cache

    Scan instances can be set to use the block cache in the RegionServer via the setCacheBlocks method. For input Scans to MapReduce jobs, this should be false. For frequently accessed rows, it is advisable to use the block cache.

    12.9.6. Optimal Loading of Row Keys

    When performing a table scan where only the row keys are needed (no families, qualifiers, values or timestamps), add a FilterList with a MUST_PASS_ALL operator to the scanner using setFilter. The filter list should include both a FirstKeyOnlyFilter and a KeyOnlyFilter. Using this filter combination will result in a worst case scenario of a RegionServer reading a single value from disk and minimal network traffic to the client for a single row.

    12.9.7. Concurrency: Monitor Data Spread

    When performing a high number of concurrent reads, monitor the data spread of the target tables. If the target table(s) have too few regions then the reads could likely be served from too few nodes.

    See Section 12.8.2, “ Table Creation: Pre-Creating Regions ”, as well as Section 12.4, “HBase Configurations”

    12.9.8. Bloom Filters

    Enabling Bloom Filters can save your having to go to disk and can help improve read latencys.

    Bloom filters were developed over in HBase-1200 Add bloomfilters.[27][28]

    See also Section 12.6.4, “Bloom Filters”. Bloom StoreFile footprint

    Bloom filters add an entry to the StoreFile general FileInfo data structure and then two extra entries to the StoreFile metadata section. BloomFilter in the StoreFile FileInfo data structure

    FileInfo has a BLOOM_FILTER_TYPE entry which is set to NONE, ROW or ROWCOL. BloomFilter entries in StoreFile metadata

    BLOOM_FILTER_META holds Bloom Size, Hash Function used, etc. Its small in size and is cached on StoreFile.Reader load

    BLOOM_FILTER_DATA is the actual bloomfilter data. Obtained on-demand. Stored in the LRU cache, if it is enabled (Its enabled by default). Bloom Filter Configuration io.hfile.bloom.enabled global kill switch

    io.hfile.bloom.enabled in Configuration serves as the kill switch in case something goes wrong. Default = true. io.hfile.bloom.error.rate

    io.hfile.bloom.error.rate = average false positive rate. Default = 1%. Decrease rate by ½ (e.g. to .5%) == +1 bit per bloom entry. io.hfile.bloom.max.fold

    io.hfile.bloom.max.fold = guaranteed minimum fold rate. Most people should leave this alone. Default = 7, or can collapse to at least 1/128th of original size. See the Development Process section of the document BloomFilters in HBase for more on what this option means.

    12.10. Deleting from HBase

    12.10.1. Using HBase Tables as Queues

    HBase tables are sometimes used as queues. In this case, special care must be taken to regularly perform major compactions on tables used in this manner. As is documented in Chapter 5, Data Model, marking rows as deleted creates additional StoreFiles which then need to be processed on reads. Tombstones only get cleaned up with major compactions.

    See also Section, “Compaction” and HBaseAdmin.majorCompact.

    12.10.2. Delete RPC Behavior

    Be aware that htable.delete(Delete) doesn't use the writeBuffer. It will execute an RegionServer RPC with each invocation. For a large number of deletes, consider htable.delete(List).


    12.11. HDFS

    Because HBase runs on Section 9.9, “HDFS” it is important to understand how it works and how it affects HBase.

    12.11.1. Current Issues With Low-Latency Reads

    The original use-case for HDFS was batch processing. As such, there low-latency reads were historically not a priority. With the increased adoption of Apache HBase this is changing, and several improvements are already in development. See the Umbrella Jira Ticket for HDFS Improvements for HBase.

    12.11.2. Leveraging local data

    Since Hadoop 1.0.0 (also 0.22.1, 0.23.1, CDH3u3 and HDP 1.0) via HDFS-2246, it is possible for the DFSClient to take a "short circuit" and read directly from disk instead of going through the DataNode when the data is local. What this means for HBase is that the RegionServers can read directly off their machine's disks instead of having to open a socket to talk to the DataNode, the former being generally much faster[29]. Also see HBase, mail # dev - read short circuit thread for more discussion around short circuit reads.

    To enable "short circuit" reads, you must set two configurations. First, the hdfs-site.xml needs to be amended. Set the property dfs.block.local-path-access.user to be the only user that can use the shortcut. This has to be the user that started HBase. Then in hbase-site.xml, set to be true

    For optimal performance when short-circuit reads are enabled, it is recommended that HDFS checksums are disabled. To maintain data integrity with HDFS checksums disabled, HBase can be configured to write its own checksums into its datablocks and verify against these. See hbase.regionserver.checksum.verify. When both local short-circuit reads and hbase level checksums are enabled, you SHOULD NOT disable configuration parameter "", which will cause skipping checksum on non-hfile reads. HBase already manages that setting under the covers.

    The DataNodes need to be restarted in order to pick up the new configuration. Be aware that if a process started under another username than the one configured here also has the shortcircuit enabled, it will get an Exception regarding an unauthorized access but the data will still be read.

    12.11.3. Performance Comparisons of HBase vs. HDFS

    A fairly common question on the dist-list is why HBase isn't as performant as HDFS files in a batch context (e.g., as a MapReduce source or sink). The short answer is that HBase is doing a lot more than HDFS (e.g., reading the KeyValues, returning the most current row or specified timestamps, etc.), and as such HBase is 4-5 times slower than HDFS in this processing context. There is room for improvement and this gap will, over time, be reduced, but HDFS will always be faster in this use-case.

    12.12. Amazon EC2

    Performance questions are common on Amazon EC2 environments because it is a shared environment. You will not see the same throughput as a dedicated server. In terms of running tests on EC2, run them several times for the same reason (i.e., it's a shared environment and you don't know what else is happening on the server).

    If you are running on EC2 and post performance questions on the dist-list, please state this fact up-front that because EC2 issues are practically a separate class of performance issues.

    12.13. Collocating HBase and MapReduce

    It is often recommended to have different clusters for HBase and MapReduce. A better qualification of this is: don't collocate a HBase that serves live requests with a heavy MR workload. OLTP and OLAP-optimized systems have conflicting requirements and one will lose to the other, usually the former. For example, short latency-sensitive disk reads will have to wait in line behind longer reads that are trying to squeeze out as much throughput as possible. MR jobs that write to HBase will also generate flushes and compactions, which will in turn invalidate blocks in the Section 9.6.4, “Block Cache”.

    If you need to process the data from your live HBase cluster in MR, you can ship the deltas with ??? or use replication to get the new data in real time on the OLAP cluster. In the worst case, if you really need to collocate both, set MR to use less Map and Reduce slots than you'd normally configure, possibly just one.

    When HBase is used for OLAP operations, it's preferable to set it up in a hardened way like configuring the ZooKeeper session timeout higher and giving more memory to the MemStores (the argument being that the Block Cache won't be used much since the workloads are usually long scans).

    12.14. Case Studies

    For Performance and Troubleshooting Case Studies, see Chapter 14, Apache HBase Case Studies.

    [26] The latest jvms do better regards fragmentation so make sure you are running a recent release. Read down in the message, Identifying concurrent mode failures caused by fragmentation.

    [27] For description of the development process -- why static blooms rather than dynamic -- and for an overview of the unique properties that pertain to blooms in HBase, as well as possible future directions, see the Development Process section of the document BloomFilters in HBase attached to HBase-1200.

    [28] The bloom filters described here are actually version two of blooms in HBase. In versions up to 0.19.x, HBase had a dynamic bloom option based on work done by the European Commission One-Lab Project 034819. The core of the HBase bloom work was later pulled up into Hadoop to implement Version 1 of HBase blooms never worked that well. Version 2 is a rewrite from scratch though again it starts with the one-lab work.

    [29] See JD's Performance Talk

    Chapter 13. Troubleshooting and Debugging Apache HBase

    Table of Contents

    13.1. General Guidelines
    13.2. Logs
    13.2.1. Log Locations
    13.2.2. Log Levels
    13.2.3. JVM Garbage Collection Logs
    13.3. Resources
    13.3.2. Mailing Lists
    13.3.3. IRC
    13.3.4. JIRA
    13.4. Tools
    13.4.1. Builtin Tools
    13.4.2. External Tools
    13.5. Client
    13.5.1. ScannerTimeoutException or UnknownScannerException
    13.5.2. LeaseException when calling
    13.5.3. Shell or client application throws lots of scary exceptions during normal operation
    13.5.4. Long Client Pauses With Compression
    13.5.5. ZooKeeper Client Connection Errors
    13.5.6. Client running out of memory though heap size seems to be stable (but the off-heap/direct heap keeps growing)
    13.5.7. Client Slowdown When Calling Admin Methods (flush, compact, etc.)
    13.5.8. Secure Client Cannot Connect ([Caused by GSSException: No valid credentials provided (Mechanism level: Failed to find any Kerberos tgt)])
    13.6. MapReduce
    13.6.1. You Think You're On The Cluster, But You're Actually Local
    13.7. NameNode
    13.7.1. HDFS Utilization of Tables and Regions
    13.7.2. Browsing HDFS for HBase Objects
    13.8. Network
    13.8.1. Network Spikes
    13.8.2. Loopback IP
    13.8.3. Network Interfaces
    13.9. RegionServer
    13.9.1. Startup Errors
    13.9.2. Runtime Errors
    13.9.3. Shutdown Errors
    13.10. Master
    13.10.1. Startup Errors
    13.10.2. Shutdown Errors
    13.11. ZooKeeper
    13.11.1. Startup Errors
    13.11.2. ZooKeeper, The Cluster Canary
    13.12. Amazon EC2
    13.12.1. ZooKeeper does not seem to work on Amazon EC2
    13.12.2. Instability on Amazon EC2
    13.12.3. Remote Java Connection into EC2 Cluster Not Working
    13.13. HBase and Hadoop version issues
    13.13.1. NoClassDefFoundError when trying to run 0.90.x on hadoop-0.20.205.x (or hadoop-1.0.x)
    13.13.2. ...cannot communicate with client version...
    13.14. Running unit or integration tests
    13.14.1. Runtime exceptions from MiniDFSCluster when running tests
    13.15. Case Studies

    13.1. General Guidelines

    Always start with the master log (TODO: Which lines?). Normally it’s just printing the same lines over and over again. If not, then there’s an issue. Google or should return some hits for those exceptions you’re seeing.

    An error rarely comes alone in Apache HBase, usually when something gets screwed up what will follow may be hundreds of exceptions and stack traces coming from all over the place. The best way to approach this type of problem is to walk the log up to where it all began, for example one trick with RegionServers is that they will print some metrics when aborting so grepping for Dump should get you around the start of the problem.

    RegionServer suicides are “normal”, as this is what they do when something goes wrong. For example, if ulimit and xcievers (the two most important initial settings, see Section, “ ulimit and nproc) aren’t changed, it will make it impossible at some point for DataNodes to create new threads that from the HBase point of view is seen as if HDFS was gone. Think about what would happen if your MySQL database was suddenly unable to access files on your local file system, well it’s the same with HBase and HDFS. Another very common reason to see RegionServers committing seppuku is when they enter prolonged garbage collection pauses that last longer than the default ZooKeeper session timeout. For more information on GC pauses, see the 3 part blog post by Todd Lipcon and Section, “Long GC pauses” above.

    13.2. Logs

    The key process logs are as follows... (replace <user> with the user that started the service, and <hostname> for the machine name)

    NameNode: $HADOOP_HOME/logs/hadoop-<user>-namenode-<hostname>.log

    DataNode: $HADOOP_HOME/logs/hadoop-<user>-datanode-<hostname>.log

    JobTracker: $HADOOP_HOME/logs/hadoop-<user>-jobtracker-<hostname>.log

    TaskTracker: $HADOOP_HOME/logs/hadoop-<user>-tasktracker-<hostname>.log

    HMaster: $HBASE_HOME/logs/hbase-<user>-master-<hostname>.log

    RegionServer: $HBASE_HOME/logs/hbase-<user>-regionserver-<hostname>.log

    ZooKeeper: TODO

    13.2.1. Log Locations

    For stand-alone deployments the logs are obviously going to be on a single machine, however this is a development configuration only. Production deployments need to run on a cluster. NameNode

    The NameNode log is on the NameNode server. The HBase Master is typically run on the NameNode server, and well as ZooKeeper.

    For smaller clusters the JobTracker is typically run on the NameNode server as well. DataNode

    Each DataNode server will have a DataNode log for HDFS, as well as a RegionServer log for HBase.

    Additionally, each DataNode server will also have a TaskTracker log for MapReduce task execution.

    13.2.2. Log Levels Enabling RPC-level logging

    Enabling the RPC-level logging on a RegionServer can often given insight on timings at the server. Once enabled, the amount of log spewed is voluminous. It is not recommended that you leave this logging on for more than short bursts of time. To enable RPC-level logging, browse to the RegionServer UI and click on Log Level. Set the log level to DEBUG for the package org.apache.hadoop.ipc (Thats right, for hadoop.ipc, NOT, hbase.ipc). Then tail the RegionServers log. Analyze.

    To disable, set the logging level back to INFO level.

    13.2.3. JVM Garbage Collection Logs

    HBase is memory intensive, and using the default GC you can see long pauses in all threads including the Juliet Pause aka "GC of Death". To help debug this or confirm this is happening GC logging can be turned on in the Java virtual machine.

    To enable, in, uncomment one of the below lines :

    # This enables basic gc logging to the .out file.
    # export SERVER_GC_OPTS="-verbose:gc -XX:+PrintGCDetails -XX:+PrintGCDateStamps"
    # This enables basic gc logging to its own file.
    # export SERVER_GC_OPTS="-verbose:gc -XX:+PrintGCDetails -XX:+PrintGCDateStamps -Xloggc:<FILE-PATH>"
    # This enables basic GC logging to its own file with automatic log rolling. Only applies to jdk 1.6.0_34+ and 1.7.0_2+.
    # export SERVER_GC_OPTS="-verbose:gc -XX:+PrintGCDetails -XX:+PrintGCDateStamps -Xloggc:<FILE-PATH> -XX:+UseGCLogFileRotation -XX:NumberOfGCLogFiles=1 -XX:GCLogFileSize=512M"
    # If <FILE-PATH> is not replaced, the log file(.gc) would be generated in the HBASE_LOG_DIR.

    At this point you should see logs like so:

    64898.952: [GC [1 CMS-initial-mark: 2811538K(3055704K)] 2812179K(3061272K), 0.0007360 secs] [Times: user=0.00 sys=0.00, real=0.00 secs]
    64898.953: [CMS-concurrent-mark-start]
    64898.971: [GC 64898.971: [ParNew: 5567K->576K(5568K), 0.0101110 secs] 2817105K->2812715K(3061272K), 0.0102200 secs] [Times: user=0.07 sys=0.00, real=0.01 secs]

    In this section, the first line indicates a 0.0007360 second pause for the CMS to initially mark. This pauses the entire VM, all threads for that period of time.

    The third line indicates a "minor GC", which pauses the VM for 0.0101110 seconds - aka 10 milliseconds. It has reduced the "ParNew" from about 5.5m to 576k. Later on in this cycle we see:

    64901.445: [CMS-concurrent-mark: 1.542/2.492 secs] [Times: user=10.49 sys=0.33, real=2.49 secs]
    64901.445: [CMS-concurrent-preclean-start]
    64901.453: [GC 64901.453: [ParNew: 5505K->573K(5568K), 0.0062440 secs] 2868746K->2864292K(3061272K), 0.0063360 secs] [Times: user=0.05 sys=0.00, real=0.01 secs]
    64901.476: [GC 64901.476: [ParNew: 5563K->575K(5568K), 0.0072510 secs] 2869283K->2864837K(3061272K), 0.0073320 secs] [Times: user=0.05 sys=0.01, real=0.01 secs]
    64901.500: [GC 64901.500: [ParNew: 5517K->573K(5568K), 0.0120390 secs] 2869780K->2865267K(3061272K), 0.0121150 secs] [Times: user=0.09 sys=0.00, real=0.01 secs]
    64901.529: [GC 64901.529: [ParNew: 5507K->569K(5568K), 0.0086240 secs] 2870200K->2865742K(3061272K), 0.0087180 secs] [Times: user=0.05 sys=0.00, real=0.01 secs]
    64901.554: [GC 64901.555: [ParNew: 5516K->575K(5568K), 0.0107130 secs] 2870689K->2866291K(3061272K), 0.0107820 secs] [Times: user=0.06 sys=0.00, real=0.01 secs]
    64901.578: [CMS-concurrent-preclean: 0.070/0.133 secs] [Times: user=0.48 sys=0.01, real=0.14 secs]
    64901.578: [CMS-concurrent-abortable-preclean-start]
    64901.584: [GC 64901.584: [ParNew: 5504K->571K(5568K), 0.0087270 secs] 2871220K->2866830K(3061272K), 0.0088220 secs] [Times: user=0.05 sys=0.00, real=0.01 secs]
    64901.609: [GC 64901.609: [ParNew: 5512K->569K(5568K), 0.0063370 secs] 2871771K->2867322K(3061272K), 0.0064230 secs] [Times: user=0.06 sys=0.00, real=0.01 secs]
    64901.615: [CMS-concurrent-abortable-preclean: 0.007/0.037 secs] [Times: user=0.13 sys=0.00, real=0.03 secs]
    64901.616: [GC[YG occupancy: 645 K (5568 K)]64901.616: [Rescan (parallel) , 0.0020210 secs]64901.618: [weak refs processing, 0.0027950 secs] [1 CMS-remark: 2866753K(3055704K)] 2867399K(3061272K), 0.0049380 secs] [Times: user=0.00 sys=0.01, real=0.01 secs]
    64901.621: [CMS-concurrent-sweep-start]

    The first line indicates that the CMS concurrent mark (finding garbage) has taken 2.4 seconds. But this is a _concurrent_ 2.4 seconds, Java has not been paused at any point in time.

    There are a few more minor GCs, then there is a pause at the 2nd last line:

    64901.616: [GC[YG occupancy: 645 K (5568 K)]64901.616: [Rescan (parallel) , 0.0020210 secs]64901.618: [weak refs processing, 0.0027950 secs] [1 CMS-remark: 2866753K(3055704K)] 2867399K(3061272K), 0.0049380 secs] [Times: user=0.00 sys=0.01, real=0.01 secs]

    The pause here is 0.0049380 seconds (aka 4.9 milliseconds) to 'remark' the heap.

    At this point the sweep starts, and you can watch the heap size go down:

    64901.637: [GC 64901.637: [ParNew: 5501K->569K(5568K), 0.0097350 secs] 2871958K->2867441K(3061272K), 0.0098370 secs] [Times: user=0.05 sys=0.00, real=0.01 secs]
    ...  lines removed ...
    64904.936: [GC 64904.936: [ParNew: 5532K->568K(5568K), 0.0070720 secs] 1365024K->1360689K(3061272K), 0.0071930 secs] [Times: user=0.05 sys=0.00, real=0.01 secs]
    64904.953: [CMS-concurrent-sweep: 2.030/3.332 secs] [Times: user=9.57 sys=0.26, real=3.33 secs]

    At this point, the CMS sweep took 3.332 seconds, and heap went from about ~ 2.8 GB to 1.3 GB (approximate).

    The key points here is to keep all these pauses low. CMS pauses are always low, but if your ParNew starts growing, you can see minor GC pauses approach 100ms, exceed 100ms and hit as high at 400ms.

    This can be due to the size of the ParNew, which should be relatively small. If your ParNew is very large after running HBase for a while, in one example a ParNew was about 150MB, then you might have to constrain the size of ParNew (The larger it is, the longer the collections take but if its too small, objects are promoted to old gen too quickly). In the below we constrain new gen size to 64m.

    Add the below line in

    export SERVER_GC_OPTS="$SERVER_GC_OPTS -XX:NewSize=64m -XX:MaxNewSize=64m"

    Similarly, to enable GC logging for client processes, uncomment one of the below lines in

    # This enables basic gc logging to the .out file.
    # export CLIENT_GC_OPTS="-verbose:gc -XX:+PrintGCDetails -XX:+PrintGCDateStamps"
    # This enables basic gc logging to its own file.
    # export CLIENT_GC_OPTS="-verbose:gc -XX:+PrintGCDetails -XX:+PrintGCDateStamps -Xloggc:<FILE-PATH>"
    # This enables basic GC logging to its own file with automatic log rolling. Only applies to jdk 1.6.0_34+ and 1.7.0_2+.
    # export CLIENT_GC_OPTS="-verbose:gc -XX:+PrintGCDetails -XX:+PrintGCDateStamps -Xloggc:<FILE-PATH> -XX:+UseGCLogFileRotation -XX:NumberOfGCLogFiles=1 -XX:GCLogFileSize=512M"
    # If <FILE-PATH> is not replaced, the log file(.gc) would be generated in the HBASE_LOG_DIR .

    For more information on GC pauses, see the 3 part blog post by Todd Lipcon and Section, “Long GC pauses” above.

    13.3. Resources

    13.3.1. indexes all the mailing lists and is great for historical searches. Search here first when you have an issue as its more than likely someone has already had your problem.

    13.3.2. Mailing Lists

    Ask a question on the Apache HBase mailing lists. The 'dev' mailing list is aimed at the community of developers actually building Apache HBase and for features currently under development, and 'user' is generally used for questions on released versions of Apache HBase. Before going to the mailing list, make sure your question has not already been answered by searching the mailing list archives first. Use Section 13.3.1, “”. Take some time crafting your question[30]; a quality question that includes all context and exhibits evidence the author has tried to find answers in the manual and out on lists is more likely to get a prompt response.

    13.3.3. IRC

    #hbase on

    13.3.4. JIRA

    JIRA is also really helpful when looking for Hadoop/HBase-specific issues.

    13.4. Tools

    13.4.1. Builtin Tools Master Web Interface

    The Master starts a web-interface on port 60010 by default.

    The Master web UI lists created tables and their definition (e.g., ColumnFamilies, blocksize, etc.). Additionally, the available RegionServers in the cluster are listed along with selected high-level metrics (requests, number of regions, usedHeap, maxHeap). The Master web UI allows navigation to each RegionServer's web UI. RegionServer Web Interface

    RegionServers starts a web-interface on port 60030 by default.

    The RegionServer web UI lists online regions and their start/end keys, as well as point-in-time RegionServer metrics (requests, regions, storeFileIndexSize, compactionQueueSize, etc.).

    See Section 15.4, “HBase Metrics” for more information in metric definitions. zkcli

    zkcli is a very useful tool for investigating ZooKeeper-related issues. To invoke:

    ./hbase zkcli -server host:port <cmd> <args>

    The commands (and arguments) are:

    	connect host:port
    	get path [watch]
    	ls path [watch]
    	set path data [version]
    	delquota [-n|-b] path
    	printwatches on|off
    	create [-s] [-e] path data acl
    	stat path [watch]
    	ls2 path [watch]
    	listquota path
    	setAcl path acl
    	getAcl path
    	sync path
    	redo cmdno
    	addauth scheme auth
    	delete path [version]
    	setquota -n|-b val path

    13.4.2. External Tools tail

    tail is the command line tool that lets you look at the end of a file. Add the “-f” option and it will refresh when new data is available. It’s useful when you are wondering what’s happening, for example, when a cluster is taking a long time to shutdown or startup as you can just fire a new terminal and tail the master log (and maybe a few RegionServers). top

    top is probably one of the most important tool when first trying to see what’s running on a machine and how the resources are consumed. Here’s an example from production system:

    top - 14:46:59 up 39 days, 11:55,  1 user,  load average: 3.75, 3.57, 3.84
    Tasks: 309 total,   1 running, 308 sleeping,   0 stopped,   0 zombie
    Cpu(s):  4.5%us,  1.6%sy,  0.0%ni, 91.7%id,  1.4%wa,  0.1%hi,  0.6%si,  0.0%st
    Mem:  24414432k total, 24296956k used,   117476k free,     7196k buffers
    Swap: 16008732k total,	14348k used, 15994384k free, 11106908k cached
    15558 hadoop	18  -2 3292m 2.4g 3556 S   79 10.4   6523:52 java
    13268 hadoop	18  -2 8967m 8.2g 4104 S   21 35.1   5170:30 java
     8895 hadoop	18  -2 1581m 497m 3420 S   11  2.1   4002:32 java

    Here we can see that the system load average during the last five minutes is 3.75, which very roughly means that on average 3.75 threads were waiting for CPU time during these 5 minutes. In general, the “perfect” utilization equals to the number of cores, under that number the machine is under utilized and over that the machine is over utilized. This is an important concept, see this article to understand it more:

    Apart from load, we can see that the system is using almost all its available RAM but most of it is used for the OS cache (which is good). The swap only has a few KBs in it and this is wanted, high numbers would indicate swapping activity which is the nemesis of performance of Java systems. Another way to detect swapping is when the load average goes through the roof (although this could also be caused by things like a dying disk, among others).

    The list of processes isn’t super useful by default, all we know is that 3 java processes are using about 111% of the CPUs. To know which is which, simply type “c” and each line will be expanded. Typing “1” will give you the detail of how each CPU is used instead of the average for all of them like shown here. jps

    jps is shipped with every JDK and gives the java process ids for the current user (if root, then it gives the ids for all users). Example:

    hadoop@sv4borg12:~$ jps
    1322 TaskTracker
    17789 HRegionServer
    27862 Child
    1158 DataNode
    25115 HQuorumPeer
    2950 Jps
    19750 ThriftServer
    18776 jmx

    In order, we see a:

    • Hadoop TaskTracker, manages the local Childs
    • HBase RegionServer, serves regions
    • Child, its MapReduce task, cannot tell which type exactly
    • Hadoop TaskTracker, manages the local Childs
    • Hadoop DataNode, serves blocks
    • HQuorumPeer, a ZooKeeper ensemble member
    • Jps, well… it’s the current process
    • ThriftServer, it’s a special one will be running only if thrift was started
    • jmx, this is a local process that’s part of our monitoring platform ( poorly named maybe). You probably don’t have that.

    You can then do stuff like checking out the full command line that started the process:

    hadoop@sv4borg12:~$ ps aux | grep HRegionServer
    hadoop   17789  155 35.2 9067824 8604364 ?     S<l  Mar04 9855:48 /usr/java/jdk1.6.0_14/bin/java -Xmx8000m -XX:+DoEscapeAnalysis -XX:+AggressiveOpts -XX:+UseConcMarkSweepGC -XX:NewSize=64m -XX:MaxNewSize=64m -XX:CMSInitiatingOccupancyFraction=88 -verbose:gc -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -Xloggc:/export1/hadoop/logs/gc-hbase.log -Dhbase.log.dir=/export1/hadoop/logs -Dhbase.log.file=hbase-hadoop-regionserver-sv4borg12.log -Dhbase.home.dir=/home/hadoop/hbase -Dhbase.root.logger=INFO,DRFA -Djava.library.path=/home/hadoop/hbase/lib/native/Linux-amd64-64 -classpath /home/hadoop/hbase/bin/../conf:[many jars]:/home/hadoop/hadoop/conf org.apache.hadoop.hbase.regionserver.HRegionServer start

    jstack is one of the most important tools when trying to figure out what a java process is doing apart from looking at the logs. It has to be used in conjunction with jps in order to give it a process id. It shows a list of threads, each one has a name, and they appear in the order that they were created (so the top ones are the most recent threads). Here’s a few example:

    The main thread of a RegionServer that’s waiting for something to do from the master:

          "regionserver60020" prio=10 tid=0x0000000040ab4000 nid=0x45cf waiting on condition [0x00007f16b6a96000..0x00007f16b6a96a70]
       java.lang.Thread.State: TIMED_WAITING (parking)
            	at sun.misc.Unsafe.park(Native Method)
            	- parking to wait for  <0x00007f16cd5c2f30> (a java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject)
            	at java.util.concurrent.locks.LockSupport.parkNanos(
            	at java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.awaitNanos(
            	at java.util.concurrent.LinkedBlockingQueue.poll(
            	The MemStore flusher thread that is currently flushing to a file:
    "regionserver60020.cacheFlusher" daemon prio=10 tid=0x0000000040f4e000 nid=0x45eb in Object.wait() [0x00007f16b5b86000..0x00007f16b5b87af0]
       java.lang.Thread.State: WAITING (on object monitor)
            	at java.lang.Object.wait(Native Method)
            	at java.lang.Object.wait(
            	- locked <0x00007f16cb14b3a8> (a org.apache.hadoop.ipc.Client$Call)
            	at org.apache.hadoop.ipc.RPC$Invoker.invoke(
            	at $Proxy1.complete(Unknown Source)
            	at sun.reflect.GeneratedMethodAccessor38.invoke(Unknown Source)
            	at sun.reflect.DelegatingMethodAccessorImpl.invoke(
            	at java.lang.reflect.Method.invoke(
            	at $Proxy1.complete(Unknown Source)
            	at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.closeInternal(
            	- locked <0x00007f16cb14b470> (a org.apache.hadoop.hdfs.DFSClient$DFSOutputStream)
            	at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.close(
            	at org.apache.hadoop.fs.FSDataOutputStream$PositionCache.close(
            	at org.apache.hadoop.fs.FSDataOutputStream.close(
            	at org.apache.hadoop.hbase.regionserver.StoreFile$Writer.close(
            	at org.apache.hadoop.hbase.regionserver.Store.internalFlushCache(
            	- locked <0x00007f16d00e6f08> (a java.lang.Object)
            	at org.apache.hadoop.hbase.regionserver.Store.flushCache(
            	at org.apache.hadoop.hbase.regionserver.Store.access$100(
            	at org.apache.hadoop.hbase.regionserver.Store$StoreFlusherImpl.flushCache(
            	at org.apache.hadoop.hbase.regionserver.HRegion.internalFlushcache(
            	at org.apache.hadoop.hbase.regionserver.HRegion.internalFlushcache(
            	at org.apache.hadoop.hbase.regionserver.HRegion.flushcache(
            	at org.apache.hadoop.hbase.regionserver.MemStoreFlusher.flushRegion(
            	at org.apache.hadoop.hbase.regionserver.MemStoreFlusher.flushRegion(

    A handler thread that’s waiting for stuff to do (like put, delete, scan, etc):

    "IPC Server handler 16 on 60020" daemon prio=10 tid=0x00007f16b011d800 nid=0x4a5e waiting on condition [0x00007f16afefd000..0x00007f16afefd9f0]
       java.lang.Thread.State: WAITING (parking)
            	at sun.misc.Unsafe.park(Native Method)
            	- parking to wait for  <0x00007f16cd3f8dd8> (a java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject)
            	at java.util.concurrent.locks.LockSupport.park(
            	at java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.await(
            	at java.util.concurrent.LinkedBlockingQueue.take(
            	at org.apache.hadoop.hbase.ipc.HBaseServer$

    And one that’s busy doing an increment of a counter (it’s in the phase where it’s trying to create a scanner in order to read the last value):

    "IPC Server handler 66 on 60020" daemon prio=10 tid=0x00007f16b006e800 nid=0x4a90 runnable [0x00007f16acb77000..0x00007f16acb77cf0]
       java.lang.Thread.State: RUNNABLE
            	at org.apache.hadoop.hbase.regionserver.KeyValueHeap.<init>(
            	at org.apache.hadoop.hbase.regionserver.StoreScanner.<init>(
            	at org.apache.hadoop.hbase.regionserver.Store.getScanner(
            	at org.apache.hadoop.hbase.regionserver.HRegion$RegionScanner.<init>(
            	at org.apache.hadoop.hbase.regionserver.HRegion.instantiateInternalScanner(
            	at org.apache.hadoop.hbase.regionserver.HRegion.getScanner(
            	at org.apache.hadoop.hbase.regionserver.HRegion.getScanner(
            	at org.apache.hadoop.hbase.regionserver.HRegion.getLastIncrement(
            	at org.apache.hadoop.hbase.regionserver.HRegion.incrementColumnValue(
            	at org.apache.hadoop.hbase.regionserver.HRegionServer.incrementColumnValue(
            	at sun.reflect.GeneratedMethodAccessor20.invoke(Unknown Source)
            	at sun.reflect.DelegatingMethodAccessorImpl.invoke(
            	at java.lang.reflect.Method.invoke(
            	at org.apache.hadoop.hbase.ipc.HBaseRPC$
            	at org.apache.hadoop.hbase.ipc.HBaseServer$

    A thread that receives data from HDFS:

    "IPC Client (47) connection to sv4borg9/ from hadoop" daemon prio=10 tid=0x00007f16a02d0000 nid=0x4fa3 runnable [0x00007f16b517d000..0x00007f16b517dbf0]
       java.lang.Thread.State: RUNNABLE
            	at Method)
            	- locked <0x00007f17d5b68c00> (a$1)
            	- locked <0x00007f17d5b68be8> (a java.util.Collections$UnmodifiableSet)
            	- locked <0x00007f1877959b50> (a
            	at org.apache.hadoop.ipc.Client$Connection$
            	- locked <0x00007f1808539178> (a
            	at org.apache.hadoop.ipc.Client$Connection.receiveResponse(
            	at org.apache.hadoop.ipc.Client$

    And here is a master trying to recover a lease after a RegionServer died:

    "LeaseChecker" daemon prio=10 tid=0x00000000407ef800 nid=0x76cd waiting on condition [0x00007f6d0eae2000..0x00007f6d0eae2a70]
       java.lang.Thread.State: WAITING (on object monitor)
            	at java.lang.Object.wait(Native Method)
            	at java.lang.Object.wait(
            	- locked <0x00007f6d1cd28f80> (a org.apache.hadoop.ipc.Client$Call)
            	at org.apache.hadoop.ipc.RPC$Invoker.invoke(
            	at $Proxy1.recoverBlock(Unknown Source)
            	at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.processDatanodeError(
            	at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.<init>(
            	at org.apache.hadoop.hdfs.DFSClient.append(
            	at org.apache.hadoop.hdfs.DistributedFileSystem.append(
            	at org.apache.hadoop.fs.FileSystem.append(
            	at org.apache.hadoop.hbase.util.FSUtils.recoverFileLease(
            	at org.apache.hadoop.hbase.regionserver.wal.HLog.splitLog(
            	at org.apache.hadoop.hbase.regionserver.wal.HLog.splitLog(
            	at org.apache.hadoop.hbase.master.HMaster.splitLogAfterStartup(
            	at org.apache.hadoop.hbase.master.HMaster.joinCluster(

    OpenTSDB is an excellent alternative to Ganglia as it uses Apache HBase to store all the time series and doesn’t have to downsample. Monitoring your own HBase cluster that hosts OpenTSDB is a good exercise.

    Here’s an example of a cluster that’s suffering from hundreds of compactions launched almost all around the same time, which severely affects the IO performance: (TODO: insert graph plotting compactionQueueSize)

    It’s a good practice to build dashboards with all the important graphs per machine and per cluster so that debugging issues can be done with a single quick look. For example, at StumbleUpon there’s one dashboard per cluster with the most important metrics from both the OS and Apache HBase. You can then go down at the machine level and get even more detailed metrics. clusterssh+top

    clusterssh+top, it’s like a poor man’s monitoring system and it can be quite useful when you have only a few machines as it’s very easy to setup. Starting clusterssh will give you one terminal per machine and another terminal in which whatever you type will be retyped in every window. This means that you can type “top” once and it will start it for all of your machines at the same time giving you full view of the current state of your cluster. You can also tail all the logs at the same time, edit files, etc.

    13.5. Client

    For more information on the HBase client, see Section 9.3, “Client”.

    13.5.1. ScannerTimeoutException or UnknownScannerException

    This is thrown if the time between RPC calls from the client to RegionServer exceeds the scan timeout. For example, if Scan.setCaching is set to 500, then there will be an RPC call to fetch the next batch of rows every 500 .next() calls on the ResultScanner because data is being transferred in blocks of 500 rows to the client. Reducing the setCaching value may be an option, but setting this value too low makes for inefficient processing on numbers of rows.

    See Section 12.9.1, “Scan Caching”.

    13.5.2. LeaseException when calling

    In some situations clients that fetch data from a RegionServer get a LeaseException instead of the usual Section 13.5.1, “ScannerTimeoutException or UnknownScannerException”. Usually the source of the exception is org.apache.hadoop.hbase.regionserver.Leases.removeLease( (line number may vary). It tends to happen in the context of a slow/freezing RegionServer#next call. It can be prevented by having hbase.rpc.timeout > Harsh J investigated the issue as part of the mailing list thread HBase, mail # user - Lease does not exist exceptions

    13.5.3. Shell or client application throws lots of scary exceptions during normal operation

    Since 0.20.0 the default log level for org.apache.hadoop.hbase.*is DEBUG.

    On your clients, edit $HBASE_HOME/conf/ and change this: to this:, or even

    13.5.4. Long Client Pauses With Compression

    This is a fairly frequent question on the Apache HBase dist-list. The scenario is that a client is typically inserting a lot of data into a relatively un-optimized HBase cluster. Compression can exacerbate the pauses, although it is not the source of the problem.

    See Section 12.8.2, “ Table Creation: Pre-Creating Regions ” on the pattern for pre-creating regions and confirm that the table isn't starting with a single region.

    See Section 12.4, “HBase Configurations” for cluster configuration, particularly hbase.hstore.blockingStoreFiles, hbase.hregion.memstore.block.multiplier, MAX_FILESIZE (region size), and MEMSTORE_FLUSHSIZE.

    A slightly longer explanation of why pauses can happen is as follows: Puts are sometimes blocked on the MemStores which are blocked by the flusher thread which is blocked because there are too many files to compact because the compactor is given too many small files to compact and has to compact the same data repeatedly. This situation can occur even with minor compactions. Compounding this situation, Apache HBase doesn't compress data in memory. Thus, the 64MB that lives in the MemStore could become a 6MB file after compression - which results in a smaller StoreFile. The upside is that more data is packed into the same region, but performance is achieved by being able to write larger files - which is why HBase waits until the flushize before writing a new StoreFile. And smaller StoreFiles become targets for compaction. Without compression the files are much bigger and don't need as much compaction, however this is at the expense of I/O.

    For additional information, see this thread on Long client pauses with compression.

    13.5.5. ZooKeeper Client Connection Errors

    Errors like this...

    11/07/05 11:26:41 WARN zookeeper.ClientCnxn: Session 0x0 for server null,
     unexpected error, closing socket connection and attempting reconnect Connection refused: no further information
            at Method)
            at Source)
            at org.apache.zookeeper.ClientCnxn$
     11/07/05 11:26:43 INFO zookeeper.ClientCnxn: Opening socket connection to
     server localhost/
     11/07/05 11:26:44 WARN zookeeper.ClientCnxn: Session 0x0 for server null,
     unexpected error, closing socket connection and attempting reconnect Connection refused: no further information
            at Method)
            at Source)
            at org.apache.zookeeper.ClientCnxn$
     11/07/05 11:26:45 INFO zookeeper.ClientCnxn: Opening socket connection to
     server localhost/

    ... are either due to ZooKeeper being down, or unreachable due to network issues.

    The utility Section, “zkcli” may help investigate ZooKeeper issues.

    13.5.6. Client running out of memory though heap size seems to be stable (but the off-heap/direct heap keeps growing)

    You are likely running into the issue that is described and worked through in the mail thread HBase, mail # user - Suspected memory leak and continued over in HBase, mail # dev - FeedbackRe: Suspected memory leak. A workaround is passing your client-side JVM a reasonable value for -XX:MaxDirectMemorySize. By default, the MaxDirectMemorySize is equal to your -Xmx max heapsize setting (if -Xmx is set). Try seting it to something smaller (for example, one user had success setting it to 1g when they had a client-side heap of 12g). If you set it too small, it will bring on FullGCs so keep it a bit hefty. You want to make this setting client-side only especially if you are running the new experiemental server-side off-heap cache since this feature depends on being able to use big direct buffers (You may have to keep separate client-side and server-side config dirs).

    13.5.7. Client Slowdown When Calling Admin Methods (flush, compact, etc.)

    This is a client issue fixed by HBASE-5073 in 0.90.6. There was a ZooKeeper leak in the client and the client was getting pummeled by ZooKeeper events with each additional invocation of the admin API.

    13.5.8. Secure Client Cannot Connect ([Caused by GSSException: No valid credentials provided (Mechanism level: Failed to find any Kerberos tgt)])

    There can be several causes that produce this symptom.

    First, check that you have a valid Kerberos ticket. One is required in order to set up communication with a secure Apache HBase cluster. Examine the ticket currently in the credential cache, if any, by running the klist command line utility. If no ticket is listed, you must obtain a ticket by running the kinit command with either a keytab specified, or by interactively entering a password for the desired principal.

    Then, consult the Java Security Guide troubleshooting section. The most common problem addressed there is resolved by setting system property value to false.

    Because of a change in the format in which MIT Kerberos writes its credentials cache, there is a bug in the Oracle JDK 6 Update 26 and earlier that causes Java to be unable to read the Kerberos credentials cache created by versions of MIT Kerberos 1.8.1 or higher. If you have this problematic combination of components in your environment, to work around this problem, first log in with kinit and then immediately refresh the credential cache with kinit -R. The refresh will rewrite the credential cache without the problematic formatting.

    Finally, depending on your Kerberos configuration, you may need to install the Java Cryptography Extension, or JCE. Insure the JCE jars are on the classpath on both server and client systems.

    You may also need to download the unlimited strength JCE policy files. Uncompress and extract the downloaded file, and install the policy jars into <java-home>/lib/security.

    13.6. MapReduce

    13.6.1. You Think You're On The Cluster, But You're Actually Local

    This following stacktrace happened using ImportTsv, but things like this can happen on any job with a mis-configuration.

        WARN mapred.LocalJobRunner: job_local_0001
    java.lang.IllegalArgumentException: Can't read partitions file
           at org.apache.hadoop.hbase.mapreduce.hadoopbackport.TotalOrderPartitioner.setConf(
           at org.apache.hadoop.util.ReflectionUtils.setConf(
           at org.apache.hadoop.util.ReflectionUtils.newInstance(
           at org.apache.hadoop.mapred.MapTask$NewOutputCollector.<init>(
           at org.apache.hadoop.mapred.MapTask.runNewMapper(
           at org.apache.hadoop.mapred.LocalJobRunner$
    Caused by: File _partition.lst does not exist.
           at org.apache.hadoop.fs.RawLocalFileSystem.getFileStatus(
           at org.apache.hadoop.fs.FilterFileSystem.getFileStatus(
           at org.apache.hadoop.fs.FileSystem.getLength(
           at org.apache.hadoop.hbase.mapreduce.hadoopbackport.TotalOrderPartitioner.readPartitions(

    .. see the critical portion of the stack? It's...

           at org.apache.hadoop.mapred.LocalJobRunner$

    LocalJobRunner means the job is running locally, not on the cluster.

    To solve this problem, you should run your MR job with your HADOOP_CLASSPATH set to include the HBase dependencies. The "hbase classpath" utility can be used to do this easily. For example (substitute VERSION with your HBase version):

              HADOOP_CLASSPATH=`hbase classpath` hadoop jar $HBASE_HOME/hbase-VERSION.jar rowcounter usertable

    See for more information on HBase MapReduce jobs and classpaths.

    13.7. NameNode

    For more information on the NameNode, see Section 9.9, “HDFS”.

    13.7.1. HDFS Utilization of Tables and Regions

    To determine how much space HBase is using on HDFS use the hadoop shell commands from the NameNode. For example...

    hadoop fs -dus /hbase/

    ...returns the summarized disk utilization for all HBase objects.

    hadoop fs -dus /hbase/myTable

    ...returns the summarized disk utilization for the HBase table 'myTable'.

    hadoop fs -du /hbase/myTable

    ...returns a list of the regions under the HBase table 'myTable' and their disk utilization.

    For more information on HDFS shell commands, see the HDFS FileSystem Shell documentation.

    13.7.2. Browsing HDFS for HBase Objects

    Somtimes it will be necessary to explore the HBase objects that exist on HDFS. These objects could include the WALs (Write Ahead Logs), tables, regions, StoreFiles, etc. The easiest way to do this is with the NameNode web application that runs on port 50070. The NameNode web application will provide links to the all the DataNodes in the cluster so that they can be browsed seamlessly.

    The HDFS directory structure of HBase tables in the cluster is...

         /<Table>             (Tables in the cluster)
              /<Region>           (Regions for the table)
                   /<ColumnFamiy>      (ColumnFamilies for the Region for the table)
                        /<StoreFile>        (StoreFiles for the ColumnFamily for the Regions for the table)

    The HDFS directory structure of HBase WAL is..

              /<RegionServer>    (RegionServers)
                   /<HLog>           (WAL HLog files for the RegionServer)

    See the HDFS User Guide for other non-shell diagnostic utilities like fsck. Zero size HLogs with data in them

    Problem: when getting a listing of all the files in a region server's .logs directory, one file has a size of 0 but it contains data.

    Answer: It's an HDFS quirk. A file that's currently being to will appear to have a size of 0 but once it's closed it will show its true size Use Cases

    Two common use-cases for querying HDFS for HBase objects is research the degree of uncompaction of a table. If there are a large number of StoreFiles for each ColumnFamily it could indicate the need for a major compaction. Additionally, after a major compaction if the resulting StoreFile is "small" it could indicate the need for a reduction of ColumnFamilies for the table.

    13.8. Network

    13.8.1. Network Spikes

    If you are seeing periodic network spikes you might want to check the compactionQueues to see if major compactions are happening.

    See Section, “Managed Compactions” for more information on managing compactions.

    13.8.2. Loopback IP

    HBase expects the loopback IP Address to be See the Getting Started section on Section, “Loopback IP”.

    13.8.3. Network Interfaces

    Are all the network interfaces functioning correctly? Are you sure? See the Troubleshooting Case Study in Section 13.15, “Case Studies”.

    13.9. RegionServer

    For more information on the RegionServers, see Section 9.6, “RegionServer”.

    13.9.1. Startup Errors Master Starts, But RegionServers Do Not

    The Master believes the RegionServers have the IP of - which is localhost and resolves to the master's own localhost.

    The RegionServers are erroneously informing the Master that their IP addresses are

    Modify /etc/hosts on the region servers, from...

    # Do not remove the following line, or various programs
    # that require network functionality will fail.               fully.qualified.regionservername regionservername  localhost.localdomain localhost
    ::1             localhost6.localdomain6 localhost6

    ... to (removing the master node's name from localhost)...

    # Do not remove the following line, or various programs
    # that require network functionality will fail.               localhost.localdomain localhost
    ::1             localhost6.localdomain6 localhost6
             Compression Link Errors

    Since compression algorithms such as LZO need to be installed and configured on each cluster this is a frequent source of startup error. If you see messages like this...

    11/02/20 01:32:15 ERROR lzo.GPLNativeCodeLoader: Could not load native gpl library
    java.lang.UnsatisfiedLinkError: no gplcompression in java.library.path
            at java.lang.ClassLoader.loadLibrary(
            at java.lang.Runtime.loadLibrary0(
            at java.lang.System.loadLibrary(

    .. then there is a path issue with the compression libraries. See the Configuration section on LZO compression configuration.

    13.9.2. Runtime Errors RegionServer Hanging

    Are you running an old JVM (< 1.6.0_u21?)? When you look at a thread dump, does it look like threads are BLOCKED but no one holds the lock all are blocked on? See HBASE 3622 Deadlock in HBaseServer (JVM bug?). Adding -XX:+UseMembar to the HBase HBASE_OPTS in conf/ may fix it.

    Also, are you using Section 9.3.4, “RowLocks”? These are discouraged because they can lock up the RegionServers if not managed properly. many open files)

    If you see log messages like this...

    2010-09-13 01:24:17,336 WARN org.apache.hadoop.hdfs.server.datanode.DataNode:
    Disk-related IOException in BlockReceiver constructor. Cause is Too many open files
            at Method)

    ... see the Getting Started section on ulimit and nproc configuration. xceiverCount 258 exceeds the limit of concurrent xcievers 256

    This typically shows up in the DataNode logs.

    See the Getting Started section on xceivers configuration. System instability, and the presence of "java.lang.OutOfMemoryError: unable to create new native thread in exceptions" HDFS DataNode logs or that of any system daemon

    See the Getting Started section on ulimit and nproc configuration. The default on recent Linux distributions is 1024 - which is far too low for HBase. DFS instability and/or RegionServer lease timeouts

    If you see warning messages like this...

    2009-02-24 10:01:33,516 WARN org.apache.hadoop.hbase.util.Sleeper: We slept xxx ms, ten times longer than scheduled: 10000
    2009-02-24 10:01:33,516 WARN org.apache.hadoop.hbase.util.Sleeper: We slept xxx ms, ten times longer than scheduled: 15000
    2009-02-24 10:01:36,472 WARN org.apache.hadoop.hbase.regionserver.HRegionServer: unable to report to master for xxx milliseconds - retrying

    ... or see full GC compactions then you may be experiencing full GC's. "No live nodes contain current block" and/or YouAreDeadException

    These errors can happen either when running out of OS file handles or in periods of severe network problems where the nodes are unreachable.

    See the Getting Started section on ulimit and nproc configuration and check your network. ZooKeeper SessionExpired events

    Master or RegionServers shutting down with messages like those in the logs:

    WARN org.apache.zookeeper.ClientCnxn: Exception
    closing session 0x278bd16a96000f to TIMED OUT
           at org.apache.zookeeper.ClientCnxn$
    WARN org.apache.hadoop.hbase.util.Sleeper: We slept 79410ms, ten times longer than scheduled: 5000
    INFO org.apache.zookeeper.ClientCnxn: Attempting connection to server hostname/IP:PORT
    INFO org.apache.zookeeper.ClientCnxn: Priming connection to java.nio.channels.SocketChannel[connected local=/IP:PORT remote=hostname/IP:PORT]
    INFO org.apache.zookeeper.ClientCnxn: Server connection successful
    WARN org.apache.zookeeper.ClientCnxn: Exception closing session 0x278bd16a96000d to Session Expired
           at org.apache.zookeeper.ClientCnxn$SendThread.readConnectResult(
           at org.apache.zookeeper.ClientCnxn$SendThread.doIO(
           at org.apache.zookeeper.ClientCnxn$
    ERROR org.apache.hadoop.hbase.regionserver.HRegionServer: ZooKeeper session expired

    The JVM is doing a long running garbage collecting which is pausing every threads (aka "stop the world"). Since the RegionServer's local ZooKeeper client cannot send heartbeats, the session times out. By design, we shut down any node that isn't able to contact the ZooKeeper ensemble after getting a timeout so that it stops serving data that may already be assigned elsewhere.

    • Make sure you give plenty of RAM (in, the default of 1GB won't be able to sustain long running imports.
    • Make sure you don't swap, the JVM never behaves well under swapping.
    • Make sure you are not CPU starving the RegionServer thread. For example, if you are running a MapReduce job using 6 CPU-intensive tasks on a machine with 4 cores, you are probably starving the RegionServer enough to create longer garbage collection pauses.
    • Increase the ZooKeeper session timeout

    If you wish to increase the session timeout, add the following to your hbase-site.xml to increase the timeout from the default of 60 seconds to 120 seconds.


    Be aware that setting a higher timeout means that the regions served by a failed RegionServer will take at least that amount of time to be transfered to another RegionServer. For a production system serving live requests, we would instead recommend setting it lower than 1 minute and over-provision your cluster in order the lower the memory load on each machines (hence having less garbage to collect per machine).

    If this is happening during an upload which only happens once (like initially loading all your data into HBase), consider bulk loading.

    See Section 13.11.2, “ZooKeeper, The Cluster Canary” for other general information about ZooKeeper troubleshooting. NotServingRegionException

    This exception is "normal" when found in the RegionServer logs at DEBUG level. This exception is returned back to the client and then the client goes back to .META. to find the new location of the moved region.

    However, if the NotServingRegionException is logged ERROR, then the client ran out of retries and something probably wrong. Regions listed by domain name, then IP

    Fix your DNS. In versions of Apache HBase before 0.92.x, reverse DNS needs to give same answer as forward lookup. See HBASE 3431 RegionServer is not using the name given it by the master; double entry in master listing of servers for gorey details. Logs flooded with '2011-01-10 12:40:48,407 INFO Got brand-new compressor' messages

    We are not using the native versions of compression libraries. See HBASE-1900 Put back native support when hadoop 0.21 is released. Copy the native libs from hadoop under hbase lib dir or symlink them into place and the message should go away. Server handler X on 60020 caught: java.nio.channels.ClosedChannelException

    If you see this type of message it means that the region server was trying to read/send data from/to a client but it already went away. Typical causes for this are if the client was killed (you see a storm of messages like this when a MapReduce job is killed or fails) or if the client receives a SocketTimeoutException. It's harmless, but you should consider digging in a bit more if you aren't doing something to trigger them.

    13.9.3. Shutdown Errors

    13.10. Master

    For more information on the Master, see Section 9.5, “Master”.

    13.10.1. Startup Errors Master says that you need to run the hbase migrations script

    Upon running that, the hbase migrations script says no files in root directory.

    HBase expects the root directory to either not exist, or to have already been initialized by hbase running a previous time. If you create a new directory for HBase using Hadoop DFS, this error will occur. Make sure the HBase root directory does not currently exist or has been initialized by a previous run of HBase. Sure fire solution is to just use Hadoop dfs to delete the HBase root and let HBase create and initialize the directory itself. Packet len6080218 is out of range!

    If you have many regions on your cluster and you see an error like that reported above in this sections title in your logs, see HBASE-4246 Cluster with too many regions cannot withstand some master failover scenarios.

    13.10.2. Shutdown Errors

    13.11. ZooKeeper

    13.11.1. Startup Errors Could not find my address: xyz in list of ZooKeeper quorum servers

    A ZooKeeper server wasn't able to start, throws that error. xyz is the name of your server.

    This is a name lookup problem. HBase tries to start a ZooKeeper server on some machine but that machine isn't able to find itself in the hbase.zookeeper.quorum configuration.

    Use the hostname presented in the error message instead of the value you used. If you have a DNS server, you can set hbase.zookeeper.dns.interface and hbase.zookeeper.dns.nameserver in hbase-site.xml to make sure it resolves to the correct FQDN.

    13.11.2. ZooKeeper, The Cluster Canary

    ZooKeeper is the cluster's "canary in the mineshaft". It'll be the first to notice issues if any so making sure its happy is the short-cut to a humming cluster.

    See the ZooKeeper Operating Environment Troubleshooting page. It has suggestions and tools for checking disk and networking performance; i.e. the operating environment your ZooKeeper and HBase are running in.

    Additionally, the utility Section, “zkcli” may help investigate ZooKeeper issues.

    13.12. Amazon EC2

    13.12.1. ZooKeeper does not seem to work on Amazon EC2

    HBase does not start when deployed as Amazon EC2 instances. Exceptions like the below appear in the Master and/or RegionServer logs:

      2009-10-19 11:52:27,030 INFO org.apache.zookeeper.ClientCnxn: Attempting
      connection to server
      2009-10-19 11:52:27,032 WARN org.apache.zookeeper.ClientCnxn: Exception
      closing session 0x0 to Connection refused

    Security group policy is blocking the ZooKeeper port on a public address. Use the internal EC2 host names when configuring the ZooKeeper quorum peer list.

    13.12.2. Instability on Amazon EC2

    Questions on HBase and Amazon EC2 come up frequently on the HBase dist-list. Search for old threads using Search Hadoop

    13.12.3. Remote Java Connection into EC2 Cluster Not Working

    See Andrew's answer here, up on the user list: Remote Java client connection into EC2 instance.

    13.13. HBase and Hadoop version issues

    13.13.1. NoClassDefFoundError when trying to run 0.90.x on hadoop-0.20.205.x (or hadoop-1.0.x)

    Apache HBase 0.90.x does not ship with hadoop-0.20.205.x, etc. To make it run, you need to replace the hadoop jars that Apache HBase shipped with in its lib directory with those of the Hadoop you want to run HBase on. If even after replacing Hadoop jars you get the below exception:

    sv4r6s38: Exception in thread "main" java.lang.NoClassDefFoundError: org/apache/commons/configuration/Configuration
    sv4r6s38:       at org.apache.hadoop.metrics2.lib.DefaultMetricsSystem.<init>(
    sv4r6s38:       at org.apache.hadoop.metrics2.lib.DefaultMetricsSystem.<clinit>(
    sv4r6s38:       at
    sv4r6s38:       at
    sv4r6s38:       at
    sv4r6s38:       at
    sv4r6s38:       at<clinit>(
    sv4r6s38:       at
    sv4r6s38:       at

    you need to copy under hbase/lib, the commons-configuration-X.jar you find in your Hadoop's lib directory. That should fix the above complaint.

    13.13.2. ...cannot communicate with client version...

    If you see something like the following in your logs ... 2012-09-24 10:20:52,168 FATAL org.apache.hadoop.hbase.master.HMaster: Unhandled exception. Starting shutdown. org.apache.hadoop.ipc.RemoteException: Server IPC version 7 cannot communicate with client version 4 ... ...are you trying to talk to an Hadoop 2.0.x from an HBase that has an Hadoop 1.0.x client? Use the HBase built against Hadoop 2.0 or rebuild your HBase passing the -Dhadoop.profile=2.0 attribute to Maven (See Section 16.8.3, “Building against various hadoop versions.” for more).

    13.14. Running unit or integration tests

    13.14.1. Runtime exceptions from MiniDFSCluster when running tests

    If you see something like the following

    java.lang.NullPointerException: null
    at org.apache.hadoop.hdfs.MiniDFSCluster.startDataNodes
    at org.apache.hadoop.hdfs.MiniDFSCluster.<init>
    at org.apache.hadoop.hbase.MiniHBaseCluster.<init>
    at org.apache.hadoop.hbase.HBaseTestingUtility.startMiniDFSCluster
    at org.apache.hadoop.hbase.HBaseTestingUtility.startMiniCluster


    ... Shutting down
    at org.apache.hadoop.hbase.MiniHBaseCluster.init
    at org.apache.hadoop.hbase.MiniHBaseCluster.<init>
    at org.apache.hadoop.hbase.MiniHBaseCluster.<init>
    at org.apache.hadoop.hbase.HBaseTestingUtility.startMiniHBaseCluster
    at org.apache.hadoop.hbase.HBaseTestingUtility.startMiniCluster

    ... then try issuing the command umask 022 before launching tests. This is a workaround for HDFS-2556

    13.15. Case Studies

    For Performance and Troubleshooting Case Studies, see Chapter 14, Apache HBase Case Studies.

    [30] See Getting Answers

    Chapter 14. Apache HBase Case Studies

    14.1. Overview

    This chapter will describe a variety of performance and troubleshooting case studies that can provide a useful blueprint on diagnosing Apache HBase cluster issues.

    For more information on Performance and Troubleshooting, see Chapter 12, Apache HBase Performance Tuning and Chapter 13, Troubleshooting and Debugging Apache HBase.

    14.2. Schema Design

    See the schema design case studies here: Section 6.11, “Schema Design Case Studies”

    14.3. Performance/Troubleshooting

    14.3.1. Case Study #1 (Performance Issue On A Single Node) Scenario

    Following a scheduled reboot, one data node began exhibiting unusual behavior. Routine MapReduce jobs run against HBase tables which regularly completed in five or six minutes began taking 30 or 40 minutes to finish. These jobs were consistently found to be waiting on map and reduce tasks assigned to the troubled data node (e.g., the slow map tasks all had the same Input Split). The situation came to a head during a distributed copy, when the copy was severely prolonged by the lagging node. Hardware


    • Two 12-core processors
    • Six Enerprise SATA disks
    • 24GB of RAM
    • Two bonded gigabit NICs


    • 10 Gigabit top-of-rack switches
    • 20 Gigabit bonded interconnects between racks. Hypotheses HBase "Hot Spot" Region

    We hypothesized that we were experiencing a familiar point of pain: a "hot spot" region in an HBase table, where uneven key-space distribution can funnel a huge number of requests to a single HBase region, bombarding the RegionServer process and cause slow response time. Examination of the HBase Master status page showed that the number of HBase requests to the troubled node was almost zero. Further, examination of the HBase logs showed that there were no region splits, compactions, or other region transitions in progress. This effectively ruled out a "hot spot" as the root cause of the observed slowness. HBase Region With Non-Local Data

    Our next hypothesis was that one of the MapReduce tasks was requesting data from HBase that was not local to the datanode, thus forcing HDFS to request data blocks from other servers over the network. Examination of the datanode logs showed that there were very few blocks being requested over the network, indicating that the HBase region was correctly assigned, and that the majority of the necessary data was located on the node. This ruled out the possibility of non-local data causing a slowdown. Excessive I/O Wait Due To Swapping Or An Over-Worked Or Failing Hard Disk

    After concluding that the Hadoop and HBase were not likely to be the culprits, we moved on to troubleshooting the datanode's hardware. Java, by design, will periodically scan its entire memory space to do garbage collection. If system memory is heavily overcommitted, the Linux kernel may enter a vicious cycle, using up all of its resources swapping Java heap back and forth from disk to RAM as Java tries to run garbage collection. Further, a failing hard disk will often retry reads and/or writes many times before giving up and returning an error. This can manifest as high iowait, as running processes wait for reads and writes to complete. Finally, a disk nearing the upper edge of its performance envelope will begin to cause iowait as it informs the kernel that it cannot accept any more data, and the kernel queues incoming data into the dirty write pool in memory. However, using vmstat(1) and free(1), we could see that no swap was being used, and the amount of disk IO was only a few kilobytes per second. Slowness Due To High Processor Usage

    Next, we checked to see whether the system was performing slowly simply due to very high computational load. top(1) showed that the system load was higher than normal, but vmstat(1) and mpstat(1) showed that the amount of processor being used for actual computation was low. Network Saturation (The Winner)

    Since neither the disks nor the processors were being utilized heavily, we moved on to the performance of the network interfaces. The datanode had two gigabit ethernet adapters, bonded to form an active-standby interface. ifconfig(8) showed some unusual anomalies, namely interface errors, overruns, framing errors. While not unheard of, these kinds of errors are exceedingly rare on modern hardware which is operating as it should:

    $ /sbin/ifconfig bond0
    bond0  Link encap:Ethernet  HWaddr 00:00:00:00:00:00  
    inet addr:10.x.x.x  Bcast:10.x.x.255  Mask:
    RX packets:2990700159 errors:12 dropped:0 overruns:1 frame:6          <--- Look Here! Errors!
    TX packets:3443518196 errors:0 dropped:0 overruns:0 carrier:0
    collisions:0 txqueuelen:0 
    RX bytes:2416328868676 (2.4 TB)  TX bytes:3464991094001 (3.4 TB)

    These errors immediately lead us to suspect that one or more of the ethernet interfaces might have negotiated the wrong line speed. This was confirmed both by running an ICMP ping from an external host and observing round-trip-time in excess of 700ms, and by running ethtool(8) on the members of the bond interface and discovering that the active interface was operating at 100Mbs/, full duplex.

    $ sudo ethtool eth0
    Settings for eth0:
    Supported ports: [ TP ]
    Supported link modes:   10baseT/Half 10baseT/Full 
                           100baseT/Half 100baseT/Full 
    Supports auto-negotiation: Yes
    Advertised link modes:  10baseT/Half 10baseT/Full 
                           100baseT/Half 100baseT/Full 
    Advertised pause frame use: No
    Advertised auto-negotiation: Yes
    Link partner advertised link modes:  Not reported
    Link partner advertised pause frame use: No
    Link partner advertised auto-negotiation: No
    Speed: 100Mb/s                                     <--- Look Here!  Should say 1000Mb/s!
    Duplex: Full
    Port: Twisted Pair
    PHYAD: 1
    Transceiver: internal
    Auto-negotiation: on
    MDI-X: Unknown
    Supports Wake-on: umbg
    Wake-on: g
    Current message level: 0x00000003 (3)
    Link detected: yes

    In normal operation, the ICMP ping round trip time should be around 20ms, and the interface speed and duplex should read, "1000MB/s", and, "Full", respectively. Resolution

    After determining that the active ethernet adapter was at the incorrect speed, we used the ifenslave(8) command to make the standby interface the active interface, which yielded an immediate improvement in MapReduce performance, and a 10 times improvement in network throughput:

    On the next trip to the datacenter, we determined that the line speed issue was ultimately caused by a bad network cable, which was replaced.

    14.3.2. Case Study #2 (Performance Research 2012)

    Investigation results of a self-described "we're not sure what's wrong, but it seems slow" problem.

    14.3.3. Case Study #3 (Performance Research 2010))

    Investigation results of general cluster performance from 2010. Although this research is on an older version of the codebase, this writeup is still very useful in terms of approach.

    14.3.4. Case Study #4 (xcievers Config)

    Case study of configuring xceivers, and diagnosing errors from mis-configurations.

    See also Section, “dfs.datanode.max.xcievers.

    Chapter 15. Apache HBase Operational Management

    This chapter will cover operational tools and practices required of a running Apache HBase cluster. The subject of operations is related to the topics of Chapter 13, Troubleshooting and Debugging Apache HBase, Chapter 12, Apache HBase Performance Tuning, and Chapter 2, Apache HBase Configuration but is a distinct topic in itself.

    15.1. HBase Tools and Utilities

    Here we list HBase tools for administration, analysis, fixup, and debugging.

    15.1.1. Health Checker

    You can configure HBase to run a script on a period and if it fails N times (configurable), have the server exit. See HBASE-7351 Periodic health check script for configurations and detail.

    15.1.2. Driver

    There is a Driver class that is executed by the HBase jar can be used to invoke frequently accessed utilities. For example,

    HADOOP_CLASSPATH=`${HBASE_HOME}/bin/hbase classpath` ${HADOOP_HOME}/bin/hadoop jar ${HBASE_HOME}/hbase-VERSION.jar

    ... will return...

    An example program must be given as the first argument.
    Valid program names are:
      completebulkload: Complete a bulk data load.
      copytable: Export a table from local cluster to peer cluster
      export: Write table data to HDFS.
      import: Import data written by Export.
      importtsv: Import data in TSV format.
      rowcounter: Count rows in HBase table
      verifyrep: Compare the data from tables in two different clusters. WARNING: It doesn't work for incrementColumnValues'd cells since the timestamp is chan

    ... for allowable program names.

    15.1.3. HBase hbck

    An fsck for your HBase install

    To run hbck against your HBase cluster run

    $ ./bin/hbase hbck

    At the end of the commands output it prints OK or INCONSISTENCY. If your cluster reports inconsistencies, pass -details to see more detail emitted. If inconsistencies, run hbck a few times because the inconsistency may be transient (e.g. cluster is starting up or a region is splitting). Passing -fix may correct the inconsistency (This latter is an experimental feature).

    For more information, see Appendix B, hbck In Depth.

    15.1.4. HFile Tool

    See Section, “HFile Tool”.

    15.1.5. WAL Tools FSHLog tool

    The main method on FSHLog offers manual split and dump facilities. Pass it WALs or the product of a split, the content of the recovered.edits. directory.

    You can get a textual dump of a WAL file content by doing the following:

     $ ./bin/hbase org.apache.hadoop.hbase.regionserver.wal.FSHLog --dump hdfs://,60020,1283516293161/ 

    The return code will be non-zero if issues with the file so you can test wholesomeness of file by redirecting STDOUT to /dev/null and testing the program return.

    Similarly you can force a split of a log file directory by doing:

     $ ./bin/hbase org.apache.hadoop.hbase.regionserver.wal.FSHLog --split hdfs://,60020,1283516293161/ HLogPrettyPrinter

    HLogPrettyPrinter is a tool with configurable options to print the contents of an HLog.

    15.1.6. Compression Tool

    See Section C.1, “CompressionTest Tool”.

    15.1.7. CopyTable

    CopyTable is a utility that can copy part or of all of a table, either to the same cluster or another cluster. The usage is as follows:

    $ bin/hbase org.apache.hadoop.hbase.mapreduce.CopyTable [--starttime=X] [--endtime=Y] [] [--peer.adr=ADR] tablename


    • starttime Beginning of the time range. Without endtime means starttime to forever.
    • endtime End of the time range. Without endtime means starttime to forever.
    • versions Number of cell versions to copy.
    • New table's name.
    • peer.adr Address of the peer cluster given in the format hbase.zookeeper.quorum:hbase.zookeeper.client.port:zookeeper.znode.parent
    • families Comma-separated list of ColumnFamilies to copy.
    • all.cells Also copy delete markers and uncollected deleted cells (advanced option).


    • tablename Name of table to copy.

    Example of copying 'TestTable' to a cluster that uses replication for a 1 hour window:

    $ bin/hbase org.apache.hadoop.hbase.mapreduce.CopyTable
    --starttime=1265875194289 --endtime=1265878794289
    --peer.adr=server1,server2,server3:2181:/hbase TestTable

    Scanner Caching

    Caching for the input Scan is configured via hbase.client.scanner.caching in the job configuration.

    See Jonathan Hsieh's Online HBase Backups with CopyTable blog post for more on CopyTable.

    15.1.8. Export

    Export is a utility that will dump the contents of table to HDFS in a sequence file. Invoke via:

    $ bin/hbase org.apache.hadoop.hbase.mapreduce.Export <tablename> <outputdir> [<versions> [<starttime> [<endtime>]]]

    Note: caching for the input Scan is configured via hbase.client.scanner.caching in the job configuration.

    15.1.9. Import

    Import is a utility that will load data that has been exported back into HBase. Invoke via:

    $ bin/hbase org.apache.hadoop.hbase.mapreduce.Import <tablename> <inputdir>

    15.1.10. ImportTsv

    ImportTsv is a utility that will load data in TSV format into HBase. It has two distinct usages: loading data from TSV format in HDFS into HBase via Puts, and preparing StoreFiles to be loaded via the completebulkload.

    To load data via Puts (i.e., non-bulk loading):

    $ bin/hbase org.apache.hadoop.hbase.mapreduce.ImportTsv -Dimporttsv.columns=a,b,c <tablename> <hdfs-inputdir>

    To generate StoreFiles for bulk-loading:

    $ bin/hbase org.apache.hadoop.hbase.mapreduce.ImportTsv -Dimporttsv.columns=a,b,c -Dimporttsv.bulk.output=hdfs://storefile-outputdir <tablename> <hdfs-data-inputdir>

    These generated StoreFiles can be loaded into HBase via Section 15.1.11, “CompleteBulkLoad”. ImportTsv Options

    Running ImportTsv with no arguments prints brief usage information:
    Usage: importtsv -Dimporttsv.columns=a,b,c <tablename> <inputdir>
    Imports the given input directory of TSV data into the specified table.
    The column names of the TSV data must be specified using the -Dimporttsv.columns
    option. This option takes the form of comma-separated column names, where each
    column name is either a simple column family, or a columnfamily:qualifier. The special
    column name HBASE_ROW_KEY is used to designate that this column should be used
    as the row key for each imported record. You must specify exactly one column
    to be the row key, and you must specify a column name for every column that exists in the
    input data.
    By default importtsv will load data directly into HBase. To instead generate
    HFiles of data to prepare for a bulk data load, pass the option:
      Note: the target table will be created with default column family descriptors if it does not already exist.
    Other options that may be specified with -D include:
      -Dimporttsv.skip.bad.lines=false - fail if encountering an invalid line
      '-Dimporttsv.separator=|' - eg separate on pipes instead of tabs
      -Dimporttsv.timestamp=currentTimeAsLong - use the specified timestamp for the import
      -Dimporttsv.mapper.class=my.Mapper - A user-defined Mapper to use instead of org.apache.hadoop.hbase.mapreduce.TsvImporterMapper
 ImportTsv Example

    For example, assume that we are loading data into a table called 'datatsv' with a ColumnFamily called 'd' with two columns "c1" and "c2".

    Assume that an input file exists as follows:

    row1	c1	c2
    row2	c1	c2
    row3	c1	c2
    row4	c1	c2
    row5	c1	c2
    row6	c1	c2
    row7	c1	c2
    row8	c1	c2
    row9	c1	c2
    row10	c1	c2

    For ImportTsv to use this imput file, the command line needs to look like this:

     HADOOP_CLASSPATH=`${HBASE_HOME}/bin/hbase classpath` ${HADOOP_HOME}/bin/hadoop jar ${HBASE_HOME}/hbase-VERSION.jar importtsv -Dimporttsv.columns=HBASE_ROW_KEY,d:c1,d:c2 -Dimporttsv.bulk.output=hdfs://storefileoutput datatsv hdfs://inputfile

    ... and in this example the first column is the rowkey, which is why the HBASE_ROW_KEY is used. The second and third columns in the file will be imported as "d:c1" and "d:c2", respectively. ImportTsv Warning

    If you have preparing a lot of data for bulk loading, make sure the target HBase table is pre-split appropriately. See Also

    For more information about bulk-loading HFiles into HBase, see Section 9.8, “Bulk Loading”

    15.1.11. CompleteBulkLoad

    The completebulkload utility will move generated StoreFiles into an HBase table. This utility is often used in conjunction with output from Section 15.1.10, “ImportTsv”.

    There are two ways to invoke this utility, with explicit classname and via the driver:

    $ bin/hbase org.apache.hadoop.hbase.mapreduce.LoadIncrementalHFiles <hdfs://storefileoutput> <tablename>

    .. and via the Driver..

    HADOOP_CLASSPATH=`${HBASE_HOME}/bin/hbase classpath` ${HADOOP_HOME}/bin/hadoop jar ${HBASE_HOME}/hbase-VERSION.jar completebulkload <hdfs://storefileoutput> <tablename>
 CompleteBulkLoad Warning

    Data generated via MapReduce is often created with file permissions that are not compatible with the running HBase process. Assuming you're running HDFS with permissions enabled, those permissions will need to be updated before you run CompleteBulkLoad.

    For more information about bulk-loading HFiles into HBase, see Section 9.8, “Bulk Loading”.

    15.1.12. WALPlayer

    WALPlayer is a utility to replay WAL files into HBase.

    The WAL can be replayed for a set of tables or all tables, and a timerange can be provided (in milliseconds). The WAL is filtered to this set of tables. The output can optionally be mapped to another set of tables.

    WALPlayer can also generate HFiles for later bulk importing, in that case only a single table and no mapping can be specified.

    Invoke via:

    $ bin/hbase org.apache.hadoop.hbase.mapreduce.WALPlayer [options] <wal inputdir> <tables> [<tableMappings>]>

    For example:

    $ bin/hbase org.apache.hadoop.hbase.mapreduce.WALPlayer /backuplogdir oldTable1,oldTable2 newTable1,newTable2

    WALPlayer, by default, runs as a mapreduce job. To NOT run WALPlayer as a mapreduce job on your cluster, force it to run all in the local process by adding the flags -Dmapred.job.tracker=local on the command line.

    15.1.13. RowCounter and CellCounter

    RowCounter is a mapreduce job to count all the rows of a table. This is a good utility to use as a sanity check to ensure that HBase can read all the blocks of a table if there are any concerns of metadata inconsistency. It will run the mapreduce all in a single process but it will run faster if you have a MapReduce cluster in place for it to exploit.

    $ bin/hbase org.apache.hadoop.hbase.mapreduce.RowCounter <tablename> [<column1> <column2>...]

    Note: caching for the input Scan is configured via hbase.client.scanner.caching in the job configuration.

    HBase ships another diagnostic mapreduce job called CellCounter. Like RowCounter, it gathers more fine-grained statistics about your table. The statistics gathered by RowCounter are more fine-grained and include:

    • Total number of rows in the table.
    • Total number of CFs across all rows.
    • Total qualifiers across all rows.
    • Total occurrence of each CF.
    • Total occurrence of each qualifier.
    • Total number of versions of each qualifier.

    The program allows you to limit the scope of the run. Provide a row regex or prefix to limit the rows to analyze. Use to specify scanning a single column family.

    $ bin/hbase org.apache.hadoop.hbase.mapreduce.CellCounter <tablename> <outputDir> [regex or prefix]

    Note: just like RowCounter, caching for the input Scan is configured via hbase.client.scanner.caching in the job configuration.

    15.1.14. mlockall

    It is possible to optionally pin your servers in physical memory making them less likely to be swapped out in oversubscribed environments by having the servers call mlockall on startup. See HBASE-4391 Add ability to start RS as root and call mlockall for how to build the optional library and have it run on startup.

    15.1.15. Offline Compaction Tool

    See the usage for the Compaction Tool. Run it like this ./bin/hbase org.apache.hadoop.hbase.regionserver.CompactionTool

    15.2. Region Management

    15.2.1. Major Compaction

    Major compactions can be requested via the HBase shell or HBaseAdmin.majorCompact.

    Note: major compactions do NOT do region merges. See Section, “Compaction” for more information about compactions.

    15.2.2. Merge

    Merge is a utility that can merge adjoining regions in the same table (see org.apache.hadoop.hbase.util.Merge).

    $ bin/hbase org.apache.hadoop.hbase.util.Merge <tablename> <region1> <region2>

    If you feel you have too many regions and want to consolidate them, Merge is the utility you need. Merge must run be done when the cluster is down. See the O'Reilly HBase Book for an example of usage.

    You will need to pass 3 parameters to this application. The first one is the table name. The second one is the fully qualified name of the first region to merge, like "table_name,\x0A,1342956111995.7cef47f192318ba7ccc75b1bbf27a82b.". The third one is the fully qualified name for the second region to merge.

    Additionally, there is a Ruby script attached to HBASE-1621 for region merging.

    15.3. Node Management

    15.3.1. Node Decommission

    You can stop an individual RegionServer by running the following script in the HBase directory on the particular node:

    $ ./bin/ stop regionserver

    The RegionServer will first close all regions and then shut itself down. On shutdown, the RegionServer's ephemeral node in ZooKeeper will expire. The master will notice the RegionServer gone and will treat it as a 'crashed' server; it will reassign the nodes the RegionServer was carrying.

    Disable the Load Balancer before Decommissioning a node

    If the load balancer runs while a node is shutting down, then there could be contention between the Load Balancer and the Master's recovery of the just decommissioned RegionServer. Avoid any problems by disabling the balancer first. See Load Balancer below.

    A downside to the above stop of a RegionServer is that regions could be offline for a good period of time. Regions are closed in order. If many regions on the server, the first region to close may not be back online until all regions close and after the master notices the RegionServer's znode gone. In Apache HBase 0.90.2, we added facility for having a node gradually shed its load and then shutdown itself down. Apache HBase 0.90.2 added the script. Here is its usage:

    $ ./bin/
    Usage: [--config &conf-dir>] [--restart] [--reload] [--thrift] [--rest] &hostname>
     thrift      If we should stop/start thrift before/after the hbase stop/start
     rest        If we should stop/start rest before/after the hbase stop/start
     restart     If we should restart after graceful stop
     reload      Move offloaded regions back on to the stopped server
     debug       Move offloaded regions back on to the stopped server
     hostname    Hostname of server we are to stop

    To decommission a loaded RegionServer, run the following:

    $ ./bin/ HOSTNAME

    where HOSTNAME is the host carrying the RegionServer you would decommission.


    The HOSTNAME passed to must match the hostname that hbase is using to identify RegionServers. Check the list of RegionServers in the master UI for how HBase is referring to servers. Its usually hostname but can also be FQDN. Whatever HBase is using, this is what you should pass the decommission script. If you pass IPs, the script is not yet smart enough to make a hostname (or FQDN) of it and so it will fail when it checks if server is currently running; the graceful unloading of regions will not run.

    The script will move the regions off the decommissioned RegionServer one at a time to minimize region churn. It will verify the region deployed in the new location before it will moves the next region and so on until the decommissioned server is carrying zero regions. At this point, the tells the RegionServer stop. The master will at this point notice the RegionServer gone but all regions will have already been redeployed and because the RegionServer went down cleanly, there will be no WAL logs to split.

    Load Balancer

    It is assumed that the Region Load Balancer is disabled while the graceful_stop script runs (otherwise the balancer and the decommission script will end up fighting over region deployments). Use the shell to disable the balancer:

    hbase(main):001:0> balance_switch false
    0 row(s) in 0.3590 seconds

    This turns the balancer OFF. To reenable, do:

    hbase(main):001:0> balance_switch true
    0 row(s) in 0.3590 seconds

    The graceful_stop will check the balancer and if enabled, will turn it off before it goes to work. If it exits prematurely because of error, it will not have reset the balancer. Hence, it is better to manage the balancer apart from graceful_stop reenabling it after you are done w/ graceful_stop. Decommissioning several Regions Servers concurrently

    If you have a large cluster, you may want to decommission more than one machine at a time by gracefully stopping mutiple RegionServers concurrently. To gracefully drain multiple regionservers at the same time, RegionServers can be put into a "draining" state. This is done by marking a RegionServer as a draining node by creating an entry in ZooKeeper under the hbase_root/draining znode. This znode has format


    just like the regionserver entries under hbase_root/rs znode.

    Without this facility, decommissioning mulitple nodes may be non-optimal because regions that are being drained from one region server may be moved to other regionservers that are also draining. Marking RegionServers to be in the draining state prevents this from happening[31]. Bad or Failing Disk

    It is good having Section, “dfs.datanode.failed.volumes.tolerated” set if you have a decent number of disks per machine for the case where a disk plain dies. But usually disks do the "John Wayne" -- i.e. take a while to go down spewing errors in dmesg -- or for some reason, run much slower than their companions. In this case you want to decommission the disk. You have two options. You can decommission the datanode or, less disruptive in that only the bad disks data will be rereplicated, can stop the datanode, unmount the bad volume (You can't umount a volume while the datanode is using it), and then restart the datanode (presuming you have set dfs.datanode.failed.volumes.tolerated > 0). The regionserver will throw some errors in its logs as it recalibrates where to get its data from -- it will likely roll its WAL log too -- but in general but for some latency spikes, it should keep on chugging.

    Short Circuit Reads

    If you are doing short-circuit reads, you will have to move the regions off the regionserver before you stop the datanode; when short-circuiting reading, though chmod'd so regionserver cannot have access, because it already has the files open, it will be able to keep reading the file blocks from the bad disk even though the datanode is down. Move the regions back after you restart the datanode.

    15.3.2. Rolling Restart

    You can also ask this script to restart a RegionServer after the shutdown AND move its old regions back into place. The latter you might do to retain data locality. A primitive rolling restart might be effected by running something like the following:

    $ for i in `cat conf/regionservers|sort`; do ./bin/ --restart --reload --debug $i; done &> /tmp/log.txt &

    Tail the output of /tmp/log.txt to follow the scripts progress. The above does RegionServers only. The script will also disable the load balancer before moving the regions. You'd need to do the master update separately. Do it before you run the above script. Here is a pseudo-script for how you might craft a rolling restart script:

    1. Untar your release, make sure of its configuration and then rsync it across the cluster. If this is 0.90.2, patch it with HBASE-3744 and HBASE-3756.

    2. Run hbck to ensure the cluster consistent

      $ ./bin/hbase hbck

      Effect repairs if inconsistent.

    3. Restart the Master:

      $ ./bin/ stop master; ./bin/ start master

    4. Run the script per RegionServer. For example:

      $ for i in `cat conf/regionservers|sort`; do ./bin/ --restart --reload --debug $i; done &> /tmp/log.txt &

      If you are running thrift or rest servers on the RegionServer, pass --thrift or --rest options (See usage for script).

    5. Restart the Master again. This will clear out dead servers list and reenable the balancer.

    6. Run hbck to ensure the cluster is consistent.

    It is important to drain HBase regions slowly when restarting regionservers. Otherwise, multiple regions go offline simultaneously as they are re-assigned to other nodes. Depending on your usage patterns, this might not be desirable.

    15.3.3. Adding a New Node

    Adding a new regionserver in HBase is essentially free, you simply start it like this:

    $ ./bin/ start regionserver

    and it will register itself with the master. Ideally you also started a DataNode on the same machine so that the RS can eventually start to have local files. If you rely on ssh to start your daemons, don't forget to add the new hostname in conf/regionservers on the master.

    At this point the region server isn't serving data because no regions have moved to it yet. If the balancer is enabled, it will start moving regions to the new RS. On a small/medium cluster this can have a very adverse effect on latency as a lot of regions will be offline at the same time. It is thus recommended to disable the balancer the same way it's done when decommissioning a node and move the regions manually (or even better, using a script that moves them one by one).

    The moved regions will all have 0% locality and won't have any blocks in cache so the region server will have to use the network to serve requests. Apart from resulting in higher latency, it may also be able to use all of your network card's capacity. For practical purposes, consider that a standard 1GigE NIC won't be able to read much more than 100MB/s. In this case, or if you are in a OLAP environment and require having locality, then it is recommended to major compact the moved regions.

    15.4. HBase Metrics

    15.4.1. Metric Setup

    See Metrics for an introduction and how to enable Metrics emission. Still valid for HBase 0.94.x.

    For HBase 0.95.x and up, see

    15.4.2. Warning To Ganglia Users

    Warning to Ganglia Users: by default, HBase will emit a LOT of metrics per RegionServer which may swamp your installation. Options include either increasing Ganglia server capacity, or configuring HBase to emit fewer metrics.

    15.4.3. Most Important RegionServer Metrics blockCacheExpressCachingRatio (formerly blockCacheHitCachingRatio)

    Block cache hit caching ratio (0 to 100). The cache-hit ratio for reads configured to look in the cache (i.e., cacheBlocks=true). callQueueLength

    Point in time length of the RegionServer call queue. If requests arrive faster than the RegionServer handlers can process them they will back up in the callQueue. compactionQueueLength (formerly compactionQueueSize)

    Point in time length of the compaction queue. This is the number of Stores in the RegionServer that have been targeted for compaction. flushQueueSize

    Point in time number of enqueued regions in the MemStore awaiting flush. hdfsBlocksLocalityIndex

    Point in time percentage of HDFS blocks that are local to this RegionServer. The higher the better. memstoreSizeMB

    Point in time sum of all the memstore sizes in this RegionServer (MB). Watch for this nearing or exceeding the configured high-watermark for MemStore memory in the RegionServer. numberOfOnlineRegions

    Point in time number of regions served by the RegionServer. This is an important metric to track for RegionServer-Region density. readRequestsCount

    Number of read requests for this RegionServer since startup. Note: this is a 32-bit integer and can roll. slowHLogAppendCount

    Number of slow HLog append writes for this RegionServer since startup, where "slow" is > 1 second. This is a good "canary" metric for HDFS. usedHeapMB

    Point in time amount of memory used by the RegionServer (MB). writeRequestsCount

    Number of write requests for this RegionServer since startup. Note: this is a 32-bit integer and can roll.

    15.4.4. Other RegionServer Metrics blockCacheCount

    Point in time block cache item count in memory. This is the number of blocks of StoreFiles (HFiles) in the cache. blockCacheEvictedCount

    Number of blocks that had to be evicted from the block cache due to heap size constraints by RegionServer since startup. blockCacheFreeMB

    Point in time block cache memory available (MB). blockCacheHitCount

    Number of blocks of StoreFiles (HFiles) read from the cache by RegionServer since startup. blockCacheHitRatio

    Block cache hit ratio (0 to 100) from RegionServer startup. Includes all read requests, although those with cacheBlocks=false will always read from disk and be counted as a "cache miss", which means that full-scan MapReduce jobs can affect this metric significantly. blockCacheMissCount

    Number of blocks of StoreFiles (HFiles) requested but not read from the cache from RegionServer startup. blockCacheSizeMB

    Point in time block cache size in memory (MB). i.e., memory in use by the BlockCache fsPreadLatency*

    There are several filesystem positional read latency (ms) metrics, all measured from RegionServer startup. fsReadLatency*

    There are several filesystem read latency (ms) metrics, all measured from RegionServer startup. The issue with interpretation is that ALL reads go into this metric (e.g., single-record Gets, full table Scans), including reads required for compactions. This metric is only interesting "over time" when comparing major releases of HBase or your own code. fsWriteLatency*

    There are several filesystem write latency (ms) metrics, all measured from RegionServer startup. The issue with interpretation is that ALL writes go into this metric (e.g., single-record Puts, full table re-writes due to compaction). This metric is only interesting "over time" when comparing major releases of HBase or your own code. NumberOfStores

    Point in time number of Stores open on the RegionServer. A Store corresponds to a ColumnFamily. For example, if a table (which contains the column family) has 3 regions on a RegionServer, there will be 3 stores open for that column family. NumberOfStorefiles

    Point in time number of StoreFiles open on the RegionServer. A store may have more than one StoreFile (HFile). requestsPerSecond

    Point in time number of read and write requests. Requests correspond to RegionServer RPC calls, thus a single Get will result in 1 request, but a Scan with caching set to 1000 will result in 1 request for each 'next' call (i.e., not each row). A bulk-load request will constitute 1 request per HFile. This metric is less interesting than readRequestsCount and writeRequestsCount in terms of measuring activity due to this metric being periodic. storeFileIndexSizeMB

    Point in time sum of all the StoreFile index sizes in this RegionServer (MB)

    15.5. HBase Monitoring

    15.5.1. Overview

    The following metrics are arguably the most important to monitor for each RegionServer for "macro monitoring", preferably with a system like OpenTSDB. If your cluster is having performance issues it's likely that you'll see something unusual with this group.



    • IO Wait
    • User CPU


    • GC

    For more information on HBase metrics, see Section 15.4, “HBase Metrics”.

    15.5.2. Slow Query Log

    The HBase slow query log consists of parseable JSON structures describing the properties of those client operations (Gets, Puts, Deletes, etc.) that either took too long to run, or produced too much output. The thresholds for "too long to run" and "too much output" are configurable, as described below. The output is produced inline in the main region server logs so that it is easy to discover further details from context with other logged events. It is also prepended with identifying tags (responseTooSlow), (responseTooLarge), (operationTooSlow), and (operationTooLarge) in order to enable easy filtering with grep, in case the user desires to see only slow queries. Configuration

    There are two configuration knobs that can be used to adjust the thresholds for when queries are logged.

    • hbase.ipc.warn.response.time Maximum number of milliseconds that a query can be run without being logged. Defaults to 10000, or 10 seconds. Can be set to -1 to disable logging by time.
    • hbase.ipc.warn.response.size Maximum byte size of response that a query can return without being logged. Defaults to 100 megabytes. Can be set to -1 to disable logging by size. Metrics

    The slow query log exposes to metrics to JMX.

    • hadoop.regionserver_rpc_slowResponse a global metric reflecting the durations of all responses that triggered logging.
    • hadoop.regionserver_rpc_methodName.aboveOneSec A metric reflecting the durations of all responses that lasted for more than one second. Output

    The output is tagged with operation e.g. (operationTooSlow) if the call was a client operation, such as a Put, Get, or Delete, which we expose detailed fingerprint information for. If not, it is tagged (responseTooSlow) and still produces parseable JSON output, but with less verbose information solely regarding its duration and size in the RPC itself. TooLarge is substituted for TooSlow if the response size triggered the logging, with TooLarge appearing even in the case that both size and duration triggered logging. Example

    2011-09-08 10:01:25,824 WARN org.apache.hadoop.ipc.HBaseServer: (operationTooSlow): {"tables":{"riley2":{"puts":[{"totalColumns":11,"families":{"actions":[{"timestamp":1315501284459,"qualifier":"0","vlen":9667580},{"timestamp":1315501284459,"qualifier":"1","vlen":10122412},{"timestamp":1315501284459,"qualifier":"2","vlen":11104617},{"timestamp":1315501284459,"qualifier":"3","vlen":13430635}]},"row":"cfcd208495d565ef66e7dff9f98764da:0"}],"families":["actions"]}},"processingtimems":956,"client":"","starttimems":1315501284456,"queuetimems":0,"totalPuts":1,"class":"HRegionServer","responsesize":0,"method":"multiPut"}

    Note that everything inside the "tables" structure is output produced by MultiPut's fingerprint, while the rest of the information is RPC-specific, such as processing time and client IP/port. Other client operations follow the same pattern and the same general structure, with necessary differences due to the nature of the individual operations. In the case that the call is not a client operation, that detailed fingerprint information will be completely absent.

    This particular example, for example, would indicate that the likely cause of slowness is simply a very large (on the order of 100MB) multiput, as we can tell by the "vlen," or value length, fields of each put in the multiPut.

    15.6. Cluster Replication

    See Cluster Replication.

    15.7. HBase Backup

    There are two broad strategies for performing HBase backups: backing up with a full cluster shutdown, and backing up on a live cluster. Each approach has pros and cons.

    For additional information, see HBase Backup Options over on the Sematext Blog.

    15.7.1. Full Shutdown Backup

    Some environments can tolerate a periodic full shutdown of their HBase cluster, for example if it is being used a back-end analytic capacity and not serving front-end web-pages. The benefits are that the NameNode/Master are RegionServers are down, so there is no chance of missing any in-flight changes to either StoreFiles or metadata. The obvious con is that the cluster is down. The steps include: Stop HBase Distcp

    Distcp could be used to either copy the contents of the HBase directory in HDFS to either the same cluster in another directory, or to a different cluster.

    Note: Distcp works in this situation because the cluster is down and there are no in-flight edits to files. Distcp-ing of files in the HBase directory is not generally recommended on a live cluster. Restore (if needed)

    The backup of the hbase directory from HDFS is copied onto the 'real' hbase directory via distcp. The act of copying these files creates new HDFS metadata, which is why a restore of the NameNode edits from the time of the HBase backup isn't required for this kind of restore, because it's a restore (via distcp) of a specific HDFS directory (i.e., the HBase part) not the entire HDFS file-system.

    15.7.2. Live Cluster Backup - Replication

    This approach assumes that there is a second cluster. See the HBase page on replication for more information.

    15.7.3. Live Cluster Backup - CopyTable

    The Section 15.1.7, “CopyTable” utility could either be used to copy data from one table to another on the same cluster, or to copy data to another table on another cluster.

    Since the cluster is up, there is a risk that edits could be missed in the copy process.

    15.7.4. Live Cluster Backup - Export

    The Section 15.1.8, “Export” approach dumps the content of a table to HDFS on the same cluster. To restore the data, the Section 15.1.9, “Import” utility would be used.

    Since the cluster is up, there is a risk that edits could be missed in the export process.

    15.8. HBase Snapshots

    HBase Snapshots allow you to take a snapshot of a table without too much impact on Region Servers. Snapshot, Clone and restore operations don't involve data copying. Also, Exporting the snapshot to another cluster doesn't have impact on the Region Servers.

    Prior to version 0.94.6, the only way to backup or to clone a table is to use CopyTable/ExportTable, or to copy all the hfiles in HDFS after disabling the table. The disadvantages of these methods are that you can degrade region server performance (Copy/Export Table) or you need to disable the table, that means no reads or writes; and this is usually unacceptable.

    15.8.1. Configuration

    To turn on the snapshot support just set the hbase.snapshot.enabled property to true. (Snapshots are enabled by default in 0.95+ and off by default in 0.94.6+)


    15.8.2. Take a Snapshot

    You can take a snapshot of a table regardless of whether it is enabled or disabled. The snapshot operation doesn't involve any data copying.

        $ ./bin/hbase shell
        hbase> snapshot 'myTable', 'myTableSnapshot-122112'

    15.8.3. Listing Snapshots

    List all snapshots taken (by printing the names and relative information).

        $ ./bin/hbase shell
        hbase> list_snapshots

    15.8.4. Deleting Snapshots

    You can remove a snapshot, and the files retained for that snapshot will be removed if no longer needed.

        $ ./bin/hbase shell
        hbase> delete_snapshot 'myTableSnapshot-122112'

    15.8.5. Clone a table from snapshot

    From a snapshot you can create a new table (clone operation) with the same data that you had when the snapshot was taken. The clone operation, doesn't involve data copies, and a change to the cloned table doesn't impact the snapshot or the original table.

        $ ./bin/hbase shell
        hbase> clone_snapshot 'myTableSnapshot-122112', 'myNewTestTable'

    15.8.6. Restore a snapshot

    The restore operation requires the table to be disabled, and the table will be restored to the state at the time when the snapshot was taken, changing both data and schema if required.

        $ ./bin/hbase shell
        hbase> disable 'myTable'
        hbase> restore_snapshot 'myTableSnapshot-122112'


    Since Replication works at log level and snapshots at file-system level, after a restore, the replicas will be in a different state from the master. If you want to use restore, you need to stop replication and redo the bootstrap.

    In case of partial data-loss due to misbehaving client, instead of a full restore that requires the table to be disabled, you can clone the table from the snapshot and use a Map-Reduce job to copy the data that you need, from the clone to the main one.

    15.8.7. Snapshots operations and ACLs

    If you are using security with the AccessController Coprocessor (See Section 8.2, “Access Control”), only a global administrator can take, clone, or restore a snapshot, and these actions do not capture the ACL rights. This means that restoring a table preserves the ACL rights of the existing table, while cloning a table creates a new table that has no ACL rights until the administrator adds them.

    15.8.8. Export to another cluster

    The ExportSnapshot tool copies all the data related to a snapshot (hfiles, logs, snapshot metadata) to another cluster. The tool executes a Map-Reduce job, similar to distcp, to copy files between the two clusters, and since it works at file-system level the hbase cluster does not have to be online.

    To copy a snapshot called MySnapshot to an HBase cluster srv2 (hdfs:///srv2:8082/hbase) using 16 mappers:

    $ bin/hbase class org.apache.hadoop.hbase.snapshot.tool.ExportSnapshot -snapshot MySnapshot -copy-to hdfs:///srv2:8082/hbase -mappers 16

    15.9. Capacity Planning and Region Sizing

    There are several considerations when planning the capacity for an HBase cluster and performing the initial configuration. Start with a solid understanding of how HBase handles data internally.

    15.9.1. Node count and hardware/VM configuration Physical data size

    Physical data size on disk is distinct from logical size of your data and is affected by the following:

    • Increased by HBase overhead
    • Decreased by compression and data block encoding, depending on data. See also this thread. You might want to test what compression and encoding (if any) make sense for your data.
    • Increased by size of region server WAL (usually fixed and negligible - less than half of RS memory size, per RS).
    • Increased by HDFS replication - usually x3.

    Aside from the disk space necessary to store the data, one RS may not be able to serve arbitrarily large amounts of data due to some practical limits on region count and size (see below). Read/Write throughput

    Number of nodes can also be driven by required thoughput for reads and/or writes. The throughput one can get per node depends a lot on data (esp. key/value sizes) and request patterns, as well as node and system configuration. Planning should be done for peak load if it is likely that the load would be the main driver of the increase of the node count. PerformanceEvaluation and YCSB tools can be used to test single node or a test cluster.

    For write, usually 5-15Mb/s per RS can be expected, since every region server has only one active WAL. There's no good estimate for reads, as it depends vastly on data, requests, and cache hit rate. Section 12.14, “Case Studies” might be helpful. JVM GC limitations

    RS cannot currently utilize very large heap due to cost of GC. There's also no good way of running multiple RS-es per server (other than running several VMs per machine). Thus, ~20-24Gb or less memory dedicated to one RS is recommended. GC tuning is required for large heap sizes. See Section, “Long GC pauses”, Section 13.2.3, “JVM Garbage Collection Logs” and elsewhere (TODO: where?)

    15.9.2. Determining region count and size

    Generally less regions makes for a smoother running cluster (you can always manually split the big regions later (if necessary) to spread the data, or request load, over the cluster); 20-200 regions per RS is a reasonable range. The number of regions cannot be configured directly (unless you go for fully manual splitting); adjust the region size to achieve the target region size given table size.

    When configuring regions for multiple tables, note that most region settings can be set on a per-table basis via HTableDescriptor, as well as shell commands. These settings will override the ones in hbase-site.xml. That is useful if your tables have different workloads/use cases.

    Also note that in the discussion of region sizes here, HDFS replication factor is not (and should not be) taken into account, whereas other factors above should be. So, if your data is compressed and replicated 3 ways by HDFS, "9 Gb region" means 9 Gb of compressed data. HDFS replication factor only affects your disk usage and is invisible to most HBase code. Number of regions per RS - upper bound

    In production scenarios, where you have a lot of data, you are normally concerned with the maximum number of regions you can have per server. Section, “Why cannot I have too many regions?” has technical discussion on the subject; in short, maximum number of regions is mostly determined by memstore memory usage. Each region has its own memstores; these grow up to a configurable size; usually in 128-256Mb range, see hbase.hregion.memstore.flush.size. There's one memstore per column family (so there's only one per region if there's one CF in the table). RS dedicates some fraction of total memory (see to region memstores. If this memory is exceeded (too much memstore usage), undesirable consequences such as unresponsive server, or later compaction storms, can result. Thus, a good starting point for the number of regions per RS (assuming one table) is

    (RS memory)*(total memstore fraction)/((memstore size)*(# column families))

    E.g. if RS has 16Gb RAM, with default settings, it is 16384*0.4/128 ~ 51 regions per RS is a starting point. The formula can be extended to multiple tables; if they all have the same configuration, just use total number of families.

    This number can be adjusted; the formula above assumes all your regions are filled at approximately the same rate. If only a fraction of your regions are going to be actively written to, you can divide the result by that fraction to get a larger region count. Then, even if all regions are written to, all region memstores are not filled evenly, and eventually jitter appears even if they are (due to limited number of concurrent flushes). Thus, one can have as many as 2-3 times more regions than the starting point; however, increased numbers carry increased risk.

    For write-heavy workload, memstore fraction can be increased in configuration at the expense of block cache; this will also allow one to have more regions. Number of regions per RS - lower bound

    HBase scales by having regions across many servers. Thus if you have 2 regions for 16GB data, on a 20 node machine your data will be concentrated on just a few machines - nearly the entire cluster will be idle. This really can't be stressed enough, since a common problem is loading 200MB data into HBase and then wondering why your awesome 10 node cluster isn't doing anything.

    On the other hand, if you have a very large amount of data, you may also want to go for a larger number of regions to avoid having regions that are too large. Maximum region size

    For large tables in production scenarios, maximum region size is mostly limited by compactions - very large compactions, esp. major, can degrade cluster performance. Currently, the recommended maximum region size is 10-20Gb, and 5-10Gb is optimal. For older 0.90.x codebase, the upper-bound of regionsize is about 4Gb, with a default of 256Mb.

    The size at which the region is split into two is generally configured via hbase.hregion.max.filesize; for details, see Section 9.7.4, “Region Splits”.

    If you cannot estimate the size of your tables well, when starting off, it's probably best to stick to the default region size, perhaps going smaller for hot tables (or manually split hot regions to spread the load over the cluster), or go with larger region sizes if your cell sizes tend to be largish (100k and up).

    In HBase 0.98, experimental stripe compactions feature was added that would allow for larger regions, especially for log data. See ???. Total data size per region server

    According to above numbers for region size and number of regions per region server, in an optimistic estimate 10 GB x 100 regions per RS will give up to 1TB served per region server, which is in line with some of the reported multi-PB use cases. However, it is important to think about the data vs cache size ratio at the RS level. With 1TB of data per server and 10 GB block cache, only 1% of the data will be cached, which may barely cover all block indices.

    15.9.3. Initial configuration and tuning

    First, see Section 2.5, “The Important Configurations”. Note that some configurations, more than others, depend on specific scenarios. Pay special attention to

    Then, there are some considerations when setting up your cluster and tables. Compactions

    Depending on read/write volume and latency requirements, optimal compaction settings may be different. See Section, “Compaction” for some details.

    When provisioning for large data sizes, however, it's good to keep in mind that compactions can affect write throughput. Thus, for write-intensive workloads, you may opt for less frequent compactions and more store files per regions. Minimum number of files for compactions (hbase.hstore.compaction.min) can be set to higher value; hbase.hstore.blockingStoreFiles should also be increased, as more files might accumulate in such case. You may also consider manually managing compactions: Section, “Managed Compactions” Pre-splitting the table

    Based on the target number of the regions per RS (see above) and number of RSes, one can pre-split the table at creation time. This would both avoid some costly splitting as the table starts to fill up, and ensure that the table starts out already distributed across many servers.

    If the table is expected to grow large enough to justify that, at least one region per RS should be created. It is not recommended to split immediately into the full target number of regions (e.g. 50 * number of RSes), but a low intermediate value can be chosen. For multiple tables, it is recommended to be conservative with presplitting (e.g. pre-split 1 region per RS at most), especially if you don't know how much each table will grow. If you split too much, you may end up with too many regions, with some tables having too many small regions.

    For pre-splitting howto, see Section 12.8.2, “ Table Creation: Pre-Creating Regions ”.

    15.10. Table Rename

    In versions 0.90.x of hbase and earlier, we had a simple script that would rename the hdfs table directory and then do an edit of the .META. table replacing all mentions of the old table name with the new. The script was called ./bin/rename_table.rb. The script was deprecated and removed mostly because it was unmaintained and the operation performed by the script was brutal.

    As of hbase 0.94.x, you can use the snapshot facility renaming a table. Here is how you would do it using the hbase shell:

    hbase shell> disable 'tableName'
    hbase shell> snapshot 'tableName', 'tableSnapshot'
    hbase shell> clone 'tableSnapshot', 'newTableName'
    hbase shell> delete_snapshot 'tableSnapshot'
    hbase shell> drop 'tableName'

    or in code it would be as follows:

    void rename(HBaseAdmin admin, String oldTableName, String newTableName) {
        String snapshotName = randomName();
        admin.snapshot(snapshotName, oldTableName);
        admin.cloneSnapshot(snapshotName, newTableName);

    [31] See this blog post for more details.

    Chapter 16. Building and Developing Apache HBase

    This chapter will be of interest only to those building and developing Apache HBase (i.e., as opposed to just downloading the latest distribution).

    16.1. Apache HBase Repositories

    There are two different repositories for Apache HBase: Subversion (SVN) and Git. The former is the system of record for committers, but the latter is easier to work with to build and contribute. SVN updates get automatically propagated to the Git repo.

    16.1.1. SVN

    svn co hbase-core-trunk

    16.1.2. Git

    git clone git://

    There is also a github repository that mirrors Apache git repository. If you'd like to develop within github environment (collaborating, pull requests) you can get the source code by:

    git clone git://

    16.2. IDEs

    16.2.1. Eclipse Code Formatting

    Under the dev-support folder, you will find hbase_eclipse_formatter.xml. We encourage you to have this formatter in place in eclipse when editing HBase code. To load it into eclipse:

    1. Go to Eclipse->Preferences...

    2. In Preferences, Go to Java->Code Style->Formatter

    3. Import... hbase_eclipse_formatter.xml

    4. Click Apply

    5. Still in Preferences, Go to Java->Editor->Save Actions

    6. Check the following:

      1. Perform the selected actions on save

      2. Format source code

      3. Format edited lines

    7. Click Apply

    In addition to the automatic formatting, make sure you follow the style guidelines explained in Section 16.11.5, “Common Patch Feedback”

    Also, no @author tags - that's a rule. Quality Javadoc comments are appreciated. And include the Apache license. Subversive Plugin

    Download and install the Subversive plugin.

    Set up an SVN Repository target from Section 16.1.1, “SVN”, then check out the code. Git Plugin

    If you cloned the project via git, download and install the Git plugin (EGit). Attach to your local git repo (via the Git Repositories window) and you'll be able to see file revision history, generate patches, etc. HBase Project Setup in Eclipse

    The easiest way is to use the m2eclipse plugin for Eclipse. Eclipse Indigo or newer has m2eclipse built-in, or it can be found here: M2Eclipse provides Maven integration for Eclipse - it even lets you use the direct Maven commands from within Eclipse to compile and test your project.

    To import the project, you merely need to go to File->Import...Maven->Existing Maven Projects and then point Eclipse at the HBase root directory; m2eclipse will automatically find all the hbase modules for you.

    If you install m2eclipse and import HBase in your workspace, you will have to fix your eclipse Build Path. Remove target folder, add target/generated-jamon and target/generated-sources/java folders. You may also remove from your Build Path the exclusions on the src/main/resources and src/test/resources to avoid error message in the console 'Failed to execute goal org.apache.maven.plugins:maven-antrun-plugin:1.6:run (default) on project hbase: 'An Ant BuildException has occured: Replace: source file .../target/classes/hbase-default.xml doesn't exist'. This will also reduce the eclipse build cycles and make your life easier when developing. Import into eclipse with the command line

    For those not inclined to use m2eclipse, you can generate the Eclipse files from the command line. First, run (you should only have to do this once):

    mvn clean install -DskipTests

    and then close Eclipse and execute...

    mvn eclipse:eclipse

    ... from your local HBase project directory in your workspace to generate some new .project and .classpathfiles. Then reopen Eclipse, or refresh your eclipse project (F5), and import the .project file in the HBase directory to a workspace. Maven Classpath Variable

    The M2_REPO classpath variable needs to be set up for the project. This needs to be set to your local Maven repository, which is usually ~/.m2/repository

    If this classpath variable is not configured, you will see compile errors in Eclipse like this...
    Description	Resource	Path	Location	Type
    The project cannot be built until build path errors are resolved	hbase		Unknown	Java Problem
    Unbound classpath variable: 'M2_REPO/asm/asm/3.1/asm-3.1.jar' in project 'hbase'	hbase		Build path	Build Path Problem
    Unbound classpath variable: 'M2_REPO/com/github/stephenc/high-scale-lib/high-scale-lib/1.1.1/high-scale-lib-1.1.1.jar' in project 'hbase'	hbase		Build path	Build Path Problem
    Unbound classpath variable: 'M2_REPO/com/google/guava/guava/r09/guava-r09.jar' in project 'hbase'	hbase		Build path	Build Path Problem
    Unbound classpath variable: 'M2_REPO/com/google/protobuf/protobuf-java/2.3.0/protobuf-java-2.3.0.jar' in project 'hbase'	hbase		Build path	Build Path Problem Unbound classpath variable:
             Eclipse Known Issues

    Eclipse will currently complain about It is not possible to turn these errors off.

    Description	Resource	Path	Location	Type
    Access restriction: The method arrayBaseOffset(Class) from the type Unsafe is not accessible due to restriction on required library /System/Library/Java/JavaVirtualMachines/1.6.0.jdk/Contents/Classes/classes.jar	/hbase/src/main/java/org/apache/hadoop/hbase/util	line 1061	Java Problem
    Access restriction: The method arrayIndexScale(Class) from the type Unsafe is not accessible due to restriction on required library /System/Library/Java/JavaVirtualMachines/1.6.0.jdk/Contents/Classes/classes.jar	/hbase/src/main/java/org/apache/hadoop/hbase/util	line 1064	Java Problem
    Access restriction: The method getLong(Object, long) from the type Unsafe is not accessible due to restriction on required library /System/Library/Java/JavaVirtualMachines/1.6.0.jdk/Contents/Classes/classes.jar	/hbase/src/main/java/org/apache/hadoop/hbase/util	line 1111	Java Problem
              Eclipse - More Information

    For additional information on setting up Eclipse for HBase development on Windows, see Michael Morello's blog on the topic.

    16.3. Building Apache HBase

    16.3.1. Basic Compile

    Thanks to maven, building HBase is pretty easy. You can read about the various maven commands in Section 16.8, “Maven Build Commands”, but the simplest command to compile HBase from its java source code is:

    mvn package -DskipTests

    Or, to clean up before compiling:

    mvn clean package -DskipTests

    With Eclipse set up as explained above in Section 16.2.1, “Eclipse”, you can also simply use the build command in Eclipse. To create the full installable HBase package takes a little bit more work, so read on.

    16.3.2. Build Protobuf

    You may need to change the protobuf definitions that reside in the hbase-protocol module or other modules.

    The protobuf files are located in hbase-protocol/src/main/protobuf. For the change to be effective, you will need to regenerate the classes. You can use maven profile compile-protobuf to do this.

    mvn compile -Dcompile-protobuf
    mvn compile -Pcompile-protobuf

    You may also want to define protoc.path for the protoc binary

    mvn compile -Dcompile-protobuf -Dprotoc.path=/opt/local/bin/protoc

    Read the hbase-protocol/README.txt for more details.

    16.3.3. Build Gotchas

    If you see Unable to find resource 'VM_global_library.vm', ignore it. Its not an error. It is officially ugly though.

    16.3.4. Building in snappy compression support

    Pass -Dsnappy to trigger the snappy maven profile for building snappy native libs into hbase. See also Section C.5, “ SNAPPY ”

    16.4. Releasing Apache HBase

    HBase 0.96.x will run on hadoop 1.x or hadoop 2.x but building, you must choose which to build against; we cannot make a single HBase binary to run against both hadoop1 and hadoop2. Since we include the Hadoop we were built against -- so we can do standalone mode -- the set of modules included in the tarball changes dependent on whether the hadoop1 or hadoop2 target chosen. You can tell which HBase you have -- whether it is for hadoop1 or hadoop2 by looking at the version; the HBase for hadoop1 will include 'hadoop1' in its version. Ditto for hadoop2.

    Maven, our build system, natively will not let you have a single product built against different dependencies. Its understandable. But neither could we convince maven to change the set of included modules and write out the correct poms w/ appropriate dependencies even though we have two build targets; one for hadoop1 and another for hadoop2. So, there is a prestep required. This prestep takes as input the current pom.xmls and it generates hadoop1 or hadoop2 versions. You then reference these generated poms when you build. Read on for examples

    Publishing to maven requires you sign the artifacts you want to upload. To have the build do this for you, you need to make sure you have a properly configured settings.xml in your local repository under .m2. Here is my ~/.m2/settings.xml.

    <settings xmlns=""
        <!- To publish a snapshot of some part of Maven -->
        <!-- To publish a website using Maven -->
        <!-- To stage a release of some part of Maven -->
        <!--Keyname is something like this ... 00A5F21E... do gpg --list-keys to find it-->

    You must use maven 3.0.x (Check by running mvn -version).

    16.4.1. Making a Release Candidate

    I'll explain by running through the process. See later in this section for more detail on particular steps. The script dev-support/ automates most of this.

    The Hadoop How To Release wiki page informs much of the below and may have more detail on particular sections so it is worth review.

    Update CHANGES.txt with the changes since the last release (query JIRA, export to excel then hack w/ vim to format to suit CHANGES.txt TODO: Needs detail). Adjust the version in all the poms appropriately; e.g. you may need to remove -SNAPSHOT from all versions. The Versions Maven Plugin can be of use here. To set a version in all poms, do something like this:

    $ mvn clean org.codehaus.mojo:versions-maven-plugin:1.3.1:set -DnewVersion=0.96.0

    Checkin the CHANGES.txt and version changes.

    Now, build the src tarball. This tarball is hadoop version independent. It is just the pure src code and documentation without an hadoop1 or hadoop2 taint. Add the -Prelease profile when building; it checks files for licenses and will fail the build if unlicensed files present.

    $ MAVEN_OPTS="-Xmx2g" mvn clean install -DskipTests assembly:single -Dassembly.file=hbase-assembly/src/main/assembly/src.xml -Prelease

    Undo the tarball and make sure it looks good (A good test is seeing if you can build from the undone tarball). Save it off to a version directory, i.e a directory somewhere where you are collecting all of the tarballs you will publish as part of the release candidate. For example if we were building a hbase-0.96.0 release candidate, we might call the directory hbase-0.96.0RC0. Later we will publish this directory as our release candidate up on

    Now we are into the making of the hadoop1 and hadoop2 specific builds. Lets do hadoop1 first. First generate the hadoop1 poms. See the script usage for what it expects by way of arguments. You will find it in the dev-support subdirectory. In the below, we generate hadoop1 poms with a version of 0.96.0-hadoop1 (the script will look for a version of 0.96.0 in the current pom.xml).

    $ ./dev-support/ 0.96.0 0.96.0-hadoop1

    The script will work silently if all goes well. It will drop a pom.xml.hadoop1 beside all pom.xmls in all modules.

    Now build the hadoop1 tarball. Note how we reference the new pom.xml.hadoop1 explicitly. We also add the -Prelease profile when building; it checks files for licenses and will fail the build if unlicensed files present. Do it in two steps. First install into the local repository and then generate documentation and assemble the tarball (Otherwise build complains that hbase modules are not in maven repo when we try to do it all in the one go especially on fresh repo). It seems that you need the install goal in both steps.

    $ MAVEN_OPTS="-Xmx3g" mvn -f pom.xml.hadoop1 clean install -DskipTests -Prelease
    $ MAVEN_OPTS="-Xmx3g" mvn -f pom.xml.hadoop1 install -DskipTests site assembly:single -Prelease

    Undo the generated tarball and check it out. Look at doc. and see if it runs, etc. Are the set of modules appropriate: e.g. do we have a hbase-hadoop2-compat in the hadoop1 tarball? If good, copy the tarball to your version directory.

    I'll tag the release at this point since its looking good. If we find an issue later, we can delete the tag and start over. Release needs to be tagged when we do next step.

    Now deploy hadoop1 hbase to mvn. Do the mvn deploy and tgz for a particular version all together in the one go else if you flip between hadoop1 and hadoop2 builds, you might mal-publish poms and hbase-default.xml's (the version interpolations won't match). This time we use the apache-release profile instead of just release profile when doing mvn deploy; it will invoke the apache pom referenced by our poms. It will also sign your artifacts published to mvn as long as your settings.xml in your local .m2 repository is configured correctly (your settings.xml adds your gpg password property to the apache profile).

    $ MAVEN_OPTS="-Xmx3g" mvn -f pom.xml.hadoop1 deploy -DskipTests -Papache-release

    The last command above copies all artifacts for hadoop1 up to mvn repo. If no -SNAPSHOT in the version, it puts the artifacts into a staging directory. This is what you want.


    See the hbase-downstreamer test for a simple example of a project that is downstream of hbase an depends on it. Check it out and run its simple test to make sure maven hbase-hadoop1 and hbase-hadoop2 are properly deployed to the maven repository.

    Lets do the hadoop2 artifacts (read above hadoop1 section closely before coming here because we don't repeat explaination in the below).

    # Generate the hadoop2 poms.
    $ ./dev-support/ 0.96.0 0.96.0-hadoop2
    # Install the hbase hadoop2 jars into local repo then build the doc and tarball
    $ MAVEN_OPTS="-Xmx3g" mvn -f pom.xml.hadoop2 clean install -DskipTests -Prelease
    $ MAVEN_OPTS="-Xmx3g" mvn -f pom.xml.hadoop2 install -DskipTests site assembly:single -Prelease
    # Undo the tgz and check it out.  If good, copy the tarball to your 'version directory'. Now deploy to mvn.
    $ MAVEN_OPTS="-Xmx3g" mvn -f pom.xml.hadoop2 deploy -DskipTests -Papache-release

    At this stage we have three tarballs in our 'version directory' and two sets of artifacts up in maven in staging area. First lets put the version directory up on You will need to sign and fingerprint them before you push them up. In the version directory do this:

    $ for i in *.tar.gz; do echo $i; gpg --print-mds $i > $i.mds ; done
    $ for i in *.tar.gz; do echo $i; gpg --armor --output $i.asc --detach-sig $i  ; done
    $ cd ..
    # Presuming our 'version directory' is named 0.96.0RC0, now copy it up to
    $ rsync -av 0.96.0RC0

    For the maven artifacts, login at Find your artifacts in the staging directory. Close the artifacts. This will give you an URL for the temporary mvn staging repository. Do the closing for hadoop1 and hadoop2 repos. See Publishing Maven Artifacts for some pointers.


    We no longer publish using the maven release plugin. Instead we do mvn deploy. It seems to give us a backdoor to maven release publishing. If no -SNAPSHOT on the version string, then we are 'deployed' to the apache maven repository staging directory from which we can publish URLs for candidates and later, if they pass, publish as release (if a -SNAPSHOT on the version string, deploy will put the artifacts up into apache snapshot repos).

    Make sure the directory is showing -- it can take a while to show -- and that the mvn repo urls are good. Announce the release candidate on the mailing list and call a vote.

    A strange issue I ran into was the one where the upload into the apache repository was being sprayed across multiple apache machines making it so I could not release. See INFRA-4482 Why is my upload to mvn spread across multiple repositories?.

    16.4.2. Publishing a SNAPSHOT to maven

    Make sure your settings.xml is set up properly (see above for how). Make sure the hbase version includes -SNAPSHOT as a suffix. Here is how I published SNAPSHOTS of a checked that had an hbase version of 0.96.0 in its poms. First we generated the hadoop1 poms with a version that has a -SNAPSHOT suffix. We then installed the build into the local repository. Then we deploy this build to apache. See the output for the location up in apache to where the snapshot is copied. Notice how add the release profile when install locally -- to find files that are without proper license -- and then the apache-release profile to deploy to the apache maven repository.

    $ ./dev-support/ 0.96.0 0.96.0-hadoop1-SNAPSHOT
     $ MAVEN_OPTS="-Xmx3g" mvn -f pom.xml.hadoop1 clean install -DskipTests  javadoc:aggregate site assembly:single -Prelease
     $ MAVEN_OPTS="-Xmx3g" mvn -f pom.xml.hadoop1 -DskipTests  deploy -Papache-release

    Next, do the same to publish the hadoop2 artifacts.

    $ ./dev-support/ 0.96.0 0.96.0-hadoop2-SNAPSHOT
    $ MAVEN_OPTS="-Xmx3g" mvn -f pom.xml.hadoop2 clean install -DskipTests  javadoc:aggregate site assembly:single -Prelease
    $ MAVEN_OPTS="-Xmx3g" mvn -f pom.xml.hadoop2 deploy -DskipTests -Papache-release

    16.5. Generating the HBase Reference Guide

    The manual is marked up using docbook. We then use the docbkx maven plugin to transform the markup to html. This plugin is run when you specify the site goal as in when you run mvn site or you can call the plugin explicitly to just generate the manual by doing mvn docbkx:generate-html (TODO: It looks like you have to run mvn site first because docbkx wants to include a transformed hbase-default.xml. Fix). When you run mvn site, we do the document generation twice, once to generate the multipage manual and then again for the single page manual (the single page version is easier to search).

    16.6. Updating

    16.6.1. Contributing to

    The Apache HBase apache web site (including this reference guide) is maintained as part of the main Apache HBase source tree, under /src/main/docbkx and /src/main/site [32]. The former -- docbkx -- is this reference guide as a bunch of xml marked up using docbook; the latter is the hbase site (the navbars, the header, the layout, etc.), and some of the documentation, legacy pages mostly that are in the process of being merged into the docbkx tree that is converted to html by a maven plugin by the site build.

    To contribute to the reference guide, edit these files under site or docbkx and submit them as a patch (see Section 16.11, “Submitting Patches”). Your Jira should contain a summary of the changes in each section (see HBASE-6081 for an example).

    To generate the site locally while you're working on it, run:

    mvn site

    Then you can load up the generated HTML files in your browser (file are under /target/site).

    16.6.2. Publishing

    As of INFRA-5680 Migrate apache hbase website, to publish the website, build it, and then deploy it over a checkout of Finally, check it in. For example, if trunk is checked out out at /Users/stack/checkouts/trunk and the hbase website,, is checked out at /Users/stack/checkouts/, to update the site, do the following:

                  # Build the site and deploy it to the checked out directory
                  # Getting the javadoc into site is a little tricky.  You have to build it before you invoke 'site'.
                  $ MAVEN_OPTS=" -Xmx3g" mvn clean install -DskipTests javadoc:aggregate site  site:stage -DstagingDirectory=/Users/stack/checkouts/

    Now check the deployed site by viewing in a brower, browse to file:////Users/stack/checkouts/ and check all is good. If all checks out, commit it and your new build will show up immediately at

                  $ cd /Users/stack/checkouts/
                  $ svn status
                  # Do an svn add of any new content...
                  $ svn add ....
                  $ svn commit -m 'Committing latest version of website...'

    16.7. Tests

    Developers, at a minimum, should familiarize themselves with the unit test detail; unit tests in HBase have a character not usually seen in other projects.

    16.7.1. Apache HBase Modules

    As of 0.96, Apache HBase is split into multiple modules which creates "interesting" rules for how and where tests are written. If you are writting code for hbase-server, see Section 16.7.2, “Unit Tests” for how to write your tests; these tests can spin up a minicluster and will need to be categorized. For any other module, for example hbase-common, the tests must be strict unit tests and just test the class under test - no use of the HBaseTestingUtility or minicluster is allowed (or even possible given the dependency tree). Running Tests in other Modules

    If the module you are developing in has no other dependencies on other HBase modules, then you can cd into that module and just run:
    mvn test
    which will just run the tests IN THAT MODULE. If there are other dependencies on other modules, then you will have run the command from the ROOT HBASE DIRECTORY. This will run the tests in the other modules, unless you specify to skip the tests in that module. For instance, to skip the tests in the hbase-server module, you would run:
    mvn clean test -PskipServerTests
    from the top level directory to run all the tests in modules other than hbase-server. Note that you can specify to skip tests in multiple modules as well as just for a single module. For example, to skip the tests in hbase-server and hbase-common, you would run:
    mvn clean test -PskipServerTests -PskipCommonTests

    Also, keep in mind that if you are running tests in the hbase-server module you will need to apply the maven profiles discussed in Section 16.7.3, “Running tests” to get the tests to run properly.

    16.7.2. Unit Tests

    Apache HBase unit tests are subdivided into four categories: small, medium, large, and integration with corresponding JUnit categories: SmallTests, MediumTests, LargeTests, IntegrationTests. JUnit categories are denoted using java annotations and look like this in your unit test code.

    public class TestHRegionInfo {
      public void testCreateHRegionInfoName() throws Exception {
        // ...

    The above example shows how to mark a unit test as belonging to the small category. All unit tests in HBase have a categorization.

    The first three categories, small, medium, and large are for tests run when you type $ mvn test; i.e. these three categorizations are for HBase unit tests. The integration category is for not for unit tests but for integration tests. These are run when you invoke $ mvn verify. Integration tests are described in Section 16.7.5, “Integration Tests” and will not be discussed further in this section on HBase unit tests.

    Apache HBase uses a patched maven surefire plugin and maven profiles to implement its unit test characterizations.

    Read the below to figure which annotation of the set small, medium, and large to put on your new HBase unit test. Small Tests

    Small tests are executed in a shared JVM. We put in this category all the tests that can be executed quickly in a shared JVM. The maximum execution time for a small test is 15 seconds, and small tests should not use a (mini)cluster. Medium Tests

    Medium tests represent tests that must be executed before proposing a patch. They are designed to run in less than 30 minutes altogether, and are quite stable in their results. They are designed to last less than 50 seconds individually. They can use a cluster, and each of them is executed in a separate JVM. Large Tests

    Large tests are everything else. They are typically large-scale tests, regression tests for specific bugs, timeout tests, performance tests. They are executed before a commit on the pre-integration machines. They can be run on the developer machine as well. Integration Tests

    Integration tests are system level tests. See Section 16.7.5, “Integration Tests” for more info.

    16.7.3. Running tests

    Below we describe how to run the Apache HBase junit categories. Default: small and medium category tests


    mvn test

    will execute all small tests in a single JVM (no fork) and then medium tests in a separate JVM for each test instance. Medium tests are NOT executed if there is an error in a small test. Large tests are NOT executed. There is one report for small tests, and one report for medium tests if they are executed. Running all tests


    mvn test -P runAllTests

    will execute small tests in a single JVM then medium and large tests in a separate JVM for each test. Medium and large tests are NOT executed if there is an error in a small test. Large tests are NOT executed if there is an error in a small or medium test. There is one report for small tests, and one report for medium and large tests if they are executed. Running a single test or all tests in a package

    To run an individual test, e.g. MyTest, do

    mvn test -Dtest=MyTest

    You can also pass multiple, individual tests as a comma-delimited list:

    mvn test -Dtest=MyTest1,MyTest2,MyTest3

    You can also pass a package, which will run all tests under the package:

    mvn test -Dtest=org.apache.hadoop.hbase.client.*

    When -Dtest is specified, localTests profile will be used. It will use the official release of maven surefire, rather than our custom surefire plugin, and the old connector (The HBase build uses a patched version of the maven surefire plugin). Each junit tests is executed in a separate JVM (A fork per test class). There is no parallelization when tests are running in this mode. You will see a new message at the end of the -report: "[INFO] Tests are skipped". It's harmless. While you need to make sure the sum of Tests run: in the Results : section of test reports matching the number of tests you specified because no error will be reported when a non-existent test case is specified. Other test invocation permutations


    mvn test -P runSmallTests

    will execute "small" tests only, using a single JVM.


    mvn test -P runMediumTests

    will execute "medium" tests only, launching a new JVM for each test-class.


    mvn test -P runLargeTests

    will execute "large" tests only, launching a new JVM for each test-class.

    For convenience, you can run

    mvn test -P runDevTests

    to execute both small and medium tests, using a single JVM. Running tests faster

    By default, $ mvn test -P runAllTests runs 5 tests in parallel. It can be increased on a developer's machine. Allowing that you can have 2 tests in parallel per core, and you need about 2Gb of memory per test (at the extreme), if you have an 8 core, 24Gb box, you can have 16 tests in parallel. but the memory available limits it to 12 (24/2), To run all tests with 12 tests in parallell, do this: mvn test -P runAllTests -Dsurefire.secondPartThreadCount=12. To increase the speed, you can as well use a ramdisk. You will need 2Gb of memory to run all tests. You will also need to delete the files between two test run. The typical way to configure a ramdisk on Linux is:

    $ sudo mkdir /ram2G
    sudo mount -t tmpfs -o size=2048M tmpfs /ram2G

    You can then use it to run all HBase tests with the command: mvn test -P runAllTests -Dsurefire.secondPartThreadCount=12

    It's also possible to use the script This script runs the medium and large tests in parallel with two maven instances, and provides a single report. This script does not use the hbase version of surefire so no parallelization is being done other than the two maven instances the script sets up. It must be executed from the directory which contains the pom.xml.

    For example running


    will execute small and medium tests. Running

    ./dev-support/ runAllTests

    will execute all tests. Running

    ./dev-support/ replayFailed

    will rerun the failed tests a second time, in a separate jvm and without parallelisation. Test Resource Checker

    A custom Maven SureFire plugin listener checks a number of resources before and after each HBase unit test runs and logs its findings at the end of the test output files which can be found in target/surefire-reports per Maven module (Tests write test reports named for the test class into this directory. Check the *-out.txt files). The resources counted are the number of threads, the number of file descriptors, etc. If the number has increased, it adds a LEAK? comment in the logs. As you can have an HBase instance running in the background, some threads can be deleted/created without any specific action in the test. However, if the test does not work as expected, or if the test should not impact these resources, it's worth checking these log lines ...hbase.ResourceChecker(157): before... and ...hbase.ResourceChecker(157): after.... For example: 2012-09-26 09:22:15,315 INFO [pool-1-thread-1] hbase.ResourceChecker(157): after: regionserver.TestColumnSeeking#testReseeking Thread=65 (was 65), OpenFileDescriptor=107 (was 107), MaxFileDescriptor=10240 (was 10240), ConnectionCount=1 (was 1)

    16.7.4. Writing Tests General rules

    • As much as possible, tests should be written as category small tests.
    • All tests must be written to support parallel execution on the same machine, hence they should not use shared resources as fixed ports or fixed file names.
    • Tests should not overlog. More than 100 lines/second makes the logs complex to read and use i/o that are hence not available for the other tests.
    • Tests can be written with HBaseTestingUtility. This class offers helper functions to create a temp directory and do the cleanup, or to start a cluster. Categories and execution time

    • All tests must be categorized, if not they could be skipped.
    • All tests should be written to be as fast as possible.
    • Small category tests should last less than 15 seconds, and must not have any side effect.
    • Medium category tests should last less than 50 seconds.
    • Large category tests should last less than 3 minutes. This should ensure a good parallelization for people using it, and ease the analysis when the test fails. Sleeps in tests

    Whenever possible, tests should not use Thread.sleep, but rather waiting for the real event they need. This is faster and clearer for the reader. Tests should not do a Thread.sleep without testing an ending condition. This allows understanding what the test is waiting for. Moreover, the test will work whatever the machine performance is. Sleep should be minimal to be as fast as possible. Waiting for a variable should be done in a 40ms sleep loop. Waiting for a socket operation should be done in a 200 ms sleep loop. Tests using a cluster

    Tests using a HRegion do not have to start a cluster: A region can use the local file system. Start/stopping a cluster cost around 10 seconds. They should not be started per test method but per test class. Started cluster must be shutdown using HBaseTestingUtility#shutdownMiniCluster, which cleans the directories. As most as possible, tests should use the default settings for the cluster. When they don't, they should document it. This will allow to share the cluster later.

    16.7.5. Integration Tests

    HBase integration/system tests are tests that are beyond HBase unit tests. They are generally long-lasting, sizeable (the test can be asked to 1M rows or 1B rows), targetable (they can take configuration that will point them at the ready-made cluster they are to run against; integration tests do not include cluster start/stop code), and verifying success, integration tests rely on public APIs only; they do not attempt to examine server internals asserting success/fail. Integration tests are what you would run when you need to more elaborate proofing of a release candidate beyond what unit tests can do. They are not generally run on the Apache Continuous Integration build server, however, some sites opt to run integration tests as a part of their continuous testing on an actual cluster.

    Integration tests currently live under the src/test directory in the hbase-it submodule and will match the regex: **/IntegrationTest*.java. All integration tests are also annotated with @Category(IntegrationTests.class).

    Integration tests can be run in two modes: using a mini cluster, or against an actual distributed cluster. Maven failsafe is used to run the tests using the mini cluster. IntegrationTestsDriver class is used for executing the tests against a distributed cluster. Integration tests SHOULD NOT assume that they are running against a mini cluster, and SHOULD NOT use private API's to access cluster state. To interact with the distributed or mini cluster uniformly, IntegrationTestingUtility, and HBaseCluster classes, and public client API's can be used.

    On a distributed cluster, integration tests that use ChaosMonkey or otherwise manipulate services thru cluster manager (e.g. restart regionservers) use SSH to do it. To run these, test process should be able to run commands on remote end, so ssh should be configured accordingly (for example, if HBase runs under hbase user in your cluster, you can set up passwordless ssh for that user and run the test also under it). To facilitate that,, and configuration settings can be used. "User" is the remote user that cluster manager should use to perform ssh commands. "Opts" contains additional options that are passed to SSH (for example, "-i /tmp/my-key"). Finally, if you have some custom environment setup, "cmd" is the override format for the entire tunnel (ssh) command. The default string is {/usr/bin/ssh %1$s %2$s%3$s%4$s "%5$s"} and is a good starting point. This is a standard Java format string with 5 arguments that is used to execute the remote command. The argument 1 (%1$s) is SSH options set the via opts setting or via environment variable, 2 is SSH user name, 3 is "@" if username is set or "" otherwise, 4 is the target host name, and 5 is the logical command to execute (that may include single quotes, so don't use them). For example, if you run the tests under non-hbase user and want to ssh as that user and change to hbase on remote machine, you can use {/usr/bin/ssh %1$s %2$s%3$s%4$s "su hbase - -c \"%5$s\""}. That way, to kill RS (for example) integration tests may run {/usr/bin/ssh some-hostname "su hbase - -c \"ps aux | ... | kill ...\""}. The command is logged in the test logs, so you can verify it is correct for your environment. Running integration tests against mini cluster

    HBase 0.92 added a verify maven target. Invoking it, for example by doing mvn verify, will run all the phases up to and including the verify phase via the maven failsafe plugin, running all the above mentioned HBase unit tests as well as tests that are in the HBase integration test group. After you have completed

    mvn install -DskipTests

    You can run just the integration tests by invoking:

    cd hbase-it
    mvn verify

    If you just want to run the integration tests in top-level, you need to run two commands. First:

    mvn failsafe:integration-test

    This actually runs ALL the integration tests.


    This command will always output BUILD SUCCESS even if there are test failures.

    At this point, you could grep the output by hand looking for failed tests. However, maven will do this for us; just use:

    mvn failsafe:verify

    The above command basically looks at all the test results (so don't remove the 'target' directory) for test failures and reports the results. Running a subset of Integration tests

    This is very similar to how you specify running a subset of unit tests (see above), but use the property it.test instead of test. To just run, use:

    mvn failsafe:integration-test -Dit.test=IntegrationTestClassXYZ

    The next thing you might want to do is run groups of integration tests, say all integration tests that are named IntegrationTestClassX*.java:

    mvn failsafe:integration-test -Dit.test=*ClassX*

    This runs everything that is an integration test that matches *ClassX*. This means anything matching: "**/IntegrationTest*ClassX*". You can also run multiple groups of integration tests using comma-delimited lists (similar to unit tests). Using a list of matches still supports full regex matching for each of the groups.This would look something like:

    mvn failsafe:integration-test -Dit.test=*ClassX*, *ClassY Running integration tests against distributed cluster

    If you have an already-setup HBase cluster, you can launch the integration tests by invoking the class IntegrationTestsDriver. You may have to run test-compile first. The configuration will be picked by the bin/hbase script.

    mvn test-compile

    Then launch the tests with:

    bin/hbase [--config config_dir] org.apache.hadoop.hbase.IntegrationTestsDriver

    Pass -h to get usage on this sweet tool. Running the IntegrationTestsDriver without any argument will launch tests found under hbase-it/src/test, having @Category(IntegrationTests.class) annotation, and a name starting with IntegrationTests. See the usage, by passing -h, to see how to filter test classes. You can pass a regex which is checked against the full class name; so, part of class name can be used. IntegrationTestsDriver uses Junit to run the tests. Currently there is no support for running integration tests against a distributed cluster using maven (see HBASE-6201).

    The tests interact with the distributed cluster by using the methods in the DistributedHBaseCluster (implementing HBaseCluster) class, which in turn uses a pluggable ClusterManager. Concrete implementations provide actual functionality for carrying out deployment-specific and environment-dependent tasks (SSH, etc). The default ClusterManager is HBaseClusterManager, which uses SSH to remotely execute start/stop/kill/signal commands, and assumes some posix commands (ps, etc). Also assumes the user running the test has enough "power" to start/stop servers on the remote machines. By default, it picks up HBASE_SSH_OPTS, HBASE_HOME, HBASE_CONF_DIR from the env, and uses bin/ to carry out the actions. Currently tarball deployments, deployments which uses, and Apache Ambari deployments are supported. /etc/init.d/ scripts are not supported for now, but it can be easily added. For other deployment options, a ClusterManager can be implemented and plugged in. Destructive integration / system tests

    In 0.96, a tool named ChaosMonkey has been introduced. It is modeled after the same-named tool by Netflix. Some of the tests use ChaosMonkey to simulate faults in the running cluster in the way of killing random servers, disconnecting servers, etc. ChaosMonkey can also be used as a stand-alone tool to run a (misbehaving) policy while you are running other tests.

    ChaosMonkey defines Action's and Policy's. Actions are sequences of events. We have at least the following actions:

    • Restart active master (sleep 5 sec)
    • Restart random regionserver (sleep 5 sec)
    • Restart random regionserver (sleep 60 sec)
    • Restart META regionserver (sleep 5 sec)
    • Restart ROOT regionserver (sleep 5 sec)
    • Batch restart of 50% of regionservers (sleep 5 sec)
    • Rolling restart of 100% of regionservers (sleep 5 sec)

    Policies on the other hand are responsible for executing the actions based on a strategy. The default policy is to execute a random action every minute based on predefined action weights. ChaosMonkey executes predefined named policies until it is stopped. More than one policy can be active at any time.

    To run ChaosMonkey as a standalone tool deploy your HBase cluster as usual. ChaosMonkey uses the configuration from the bin/hbase script, thus no extra configuration needs to be done. You can invoke the ChaosMonkey by running:

    bin/hbase org.apache.hadoop.hbase.util.ChaosMonkey

    This will output smt like:

    12/11/19 23:21:57 INFO util.ChaosMonkey: Using ChaosMonkey Policy: class org.apache.hadoop.hbase.util.ChaosMonkey$PeriodicRandomActionPolicy, period:60000
    12/11/19 23:21:57 INFO util.ChaosMonkey: Sleeping for 26953 to add jitter
    12/11/19 23:22:24 INFO util.ChaosMonkey: Performing action: Restart active master
    12/11/19 23:22:24 INFO util.ChaosMonkey: Killing,60000,1353367210440
    12/11/19 23:22:24 INFO hbase.HBaseCluster: Aborting Master:,60000,1353367210440
    12/11/19 23:22:24 INFO hbase.ClusterManager: Executing remote command: ps aux | grep master | grep -v grep | tr -s ' ' | cut -d ' ' -f2 | xargs kill -s SIGKILL ,
    12/11/19 23:22:25 INFO hbase.ClusterManager: Executed remote command, exit code:0 , output:
    12/11/19 23:22:25 INFO hbase.HBaseCluster: Waiting service:master to stop:,60000,1353367210440
    12/11/19 23:22:25 INFO hbase.ClusterManager: Executing remote command: ps aux | grep master | grep -v grep | tr -s ' ' | cut -d ' ' -f2 ,
    12/11/19 23:22:25 INFO hbase.ClusterManager: Executed remote command, exit code:0 , output:
    12/11/19 23:22:25 INFO util.ChaosMonkey: Killed master,60000,1353367210440
    12/11/19 23:22:25 INFO util.ChaosMonkey: Sleeping for:5000
    12/11/19 23:22:30 INFO util.ChaosMonkey: Starting
    12/11/19 23:22:30 INFO hbase.HBaseCluster: Starting Master on:
    12/11/19 23:22:30 INFO hbase.ClusterManager: Executing remote command: /homes/enis/code/hbase-0.94/bin/../bin/ --config /homes/enis/code/hbase-0.94/bin/../conf start master ,
    12/11/19 23:22:31 INFO hbase.ClusterManager: Executed remote command, exit code:0 , output:starting master, logging to /homes/enis/code/hbase-0.94/bin/../logs/
    12/11/19 23:22:33 INFO util.ChaosMonkey: Started master:,60000,1353367210440
    12/11/19 23:22:33 INFO util.ChaosMonkey: Sleeping for:51321
    12/11/19 23:23:24 INFO util.ChaosMonkey: Performing action: Restart random region server
    12/11/19 23:23:24 INFO util.ChaosMonkey: Killing region,60020,1353367027826
    12/11/19 23:23:24 INFO hbase.HBaseCluster: Aborting RS:,60020,1353367027826
    12/11/19 23:23:24 INFO hbase.ClusterManager: Executing remote command: ps aux | grep regionserver | grep -v grep | tr -s ' ' | cut -d ' ' -f2 | xargs kill -s SIGKILL ,
    12/11/19 23:23:25 INFO hbase.ClusterManager: Executed remote command, exit code:0 , output:
    12/11/19 23:23:25 INFO hbase.HBaseCluster: Waiting service:regionserver to stop:,60020,1353367027826
    12/11/19 23:23:25 INFO hbase.ClusterManager: Executing remote command: ps aux | grep regionserver | grep -v grep | tr -s ' ' | cut -d ' ' -f2 ,
    12/11/19 23:23:25 INFO hbase.ClusterManager: Executed remote command, exit code:0 , output:
    12/11/19 23:23:25 INFO util.ChaosMonkey: Killed region,60020,1353367027826. Reported num of rs:6
    12/11/19 23:23:25 INFO util.ChaosMonkey: Sleeping for:60000
    12/11/19 23:24:25 INFO util.ChaosMonkey: Starting region
    12/11/19 23:24:25 INFO hbase.HBaseCluster: Starting RS on:
    12/11/19 23:24:25 INFO hbase.ClusterManager: Executing remote command: /homes/enis/code/hbase-0.94/bin/../bin/ --config /homes/enis/code/hbase-0.94/bin/../conf start regionserver ,
    12/11/19 23:24:26 INFO hbase.ClusterManager: Executed remote command, exit code:0 , output:starting regionserver, logging to /homes/enis/code/hbase-0.94/bin/../logs/
    12/11/19 23:24:27 INFO util.ChaosMonkey: Started region,60020,1353367027826. Reported num of rs:6

    As you can see from the log, ChaosMonkey started the default PeriodicRandomActionPolicy, which is configured with all the available actions, and ran RestartActiveMaster and RestartRandomRs actions. ChaosMonkey tool, if run from command line, will keep on running until the process is killed.

    16.8. Maven Build Commands

    All commands executed from the local HBase project directory.

    Note: use Maven 3 (Maven 2 may work but we suggest you use Maven 3).

    16.8.1. Compile

    mvn compile

    16.8.2. Running all or individual Unit Tests

    See the Section 16.7.3, “Running tests” section above in Section 16.7.2, “Unit Tests”

    16.8.3. Building against various hadoop versions.

    As of 0.96, Apache HBase supports building against Apache Hadoop versions: 1.0.3, 2.0.0-alpha and 3.0.0-SNAPSHOT. By default, we will build with Hadoop-1.0.3. To change the version to run with Hadoop-2.0.0-alpha, you would run:

    mvn -Dhadoop.profile=2.0 ...

    That is, designate build with hadoop.profile 2.0. Pass 2.0 for hadoop.profile to build against hadoop 2.0. Tests may not all pass as of this writing so you may need to pass -DskipTests unless you are inclined to fix the failing tests.

    Similarly, for 3.0, you would just replace the profile value. Note that Hadoop-3.0.0-SNAPSHOT does not currently have a deployed maven artificat - you will need to build and install your own in your local maven repository if you want to run against this profile.

    In earilier verions of Apache HBase, you can build against older versions of Apache Hadoop, notably, Hadoop 0.22.x and 0.23.x. If you are running, for example HBase-0.94 and wanted to build against Hadoop 0.23.x, you would run with:

    mvn -Dhadoop.profile=22 ...

    16.9. Getting Involved

    Apache HBase gets better only when people contribute!

    As Apache HBase is an Apache Software Foundation project, see Appendix H, HBase and the Apache Software Foundation for more information about how the ASF functions.

    16.9.1. Mailing Lists

    Sign up for the dev-list and the user-list. See the mailing lists page. Posing questions - and helping to answer other people's questions - is encouraged! There are varying levels of experience on both lists so patience and politeness are encouraged (and please stay on topic.)

    16.9.2. Jira

    Check for existing issues in Jira. If it's either a new feature request, enhancement, or a bug, file a ticket. Jira Priorities

    The following is a guideline on setting Jira issue priorities:

    • Blocker: Should only be used if the issue WILL cause data loss or cluster instability reliably.
    • Critical: The issue described can cause data loss or cluster instability in some cases.
    • Major: Important but not tragic issues, like updates to the client API that will add a lot of much-needed functionality or significant bugs that need to be fixed but that don't cause data loss.
    • Minor: Useful enhancements and annoying but not damaging bugs.
    • Trivial: Useful enhancements but generally cosmetic. Code Blocks in Jira Comments

    A commonly used macro in Jira is {code}. If you do this in a Jira comment...

       code snippet

    ... Jira will format the code snippet like code, instead of a regular comment. It improves readability.

    16.10. Developing

    16.10.1. Codelines

    Most development is done on TRUNK. However, there are branches for minor releases (e.g., 0.90.1, 0.90.2, and 0.90.3 are on the 0.90 branch).

    If you have any questions on this just send an email to the dev dist-list.

    16.10.2. Unit Tests

    In HBase we use JUnit 4. If you need to run miniclusters of HDFS, ZooKeeper, HBase, or MapReduce testing, be sure to checkout the HBaseTestingUtility. Alex Baranau of Sematext describes how it can be used in HBase Case-Study: Using HBaseTestingUtility for Local Testing and Development (2010). Mockito

    Sometimes you don't need a full running server unit testing. For example, some methods can make do with a a org.apache.hadoop.hbase.Server instance or a org.apache.hadoop.hbase.master.MasterServices Interface reference rather than a full-blown org.apache.hadoop.hbase.master.HMaster. In these cases, you maybe able to get away with a mocked Server instance. For example:


    16.10.3. Code Standards

    See Section, “Code Formatting” and Section 16.11.5, “Common Patch Feedback”.

    Also, please pay attention to the interface stability/audience classifications that you will see all over our code base. They look like this at the head of the class:


    If the InterfaceAudience is Private, we can change the class (and we do not need to include a InterfaceStability mark). If a class is marked Public but its InterfaceStability is marked Unstable, we can change it. If it's marked Public/Evolving, we're allowed to change it but should try not to. If it's Public and Stable we can't change it without a deprecation path or with a really GREAT reason.

    When you add new classes, mark them with the annotations above if publically accessible. If you are not cleared on how to mark your additions, ask up on the dev list.

    This convention comes from our parent project Hadoop.

    16.10.4. Invariants

    We don't have many but what we have we list below. All are subject to challenge of course but until then, please hold to the rules of the road. No permanent state in ZooKeeper

    ZooKeeper state should transient (treat it like memory). If deleted, hbase should be able to recover and essentially be in the same state[33].

    16.10.5. Running In-Situ

    If you are developing Apache HBase, frequently it is useful to test your changes against a more-real cluster than what you find in unit tests. In this case, HBase can be run directly from the source in local-mode. All you need to do is run:


    This will spin up a full local-cluster, just as if you had packaged up HBase and installed it on your machine.

    Keep in mind that you will need to have installed HBase into your local maven repository for the in-situ cluster to work properly. That is, you will need to run:

    mvn clean install -DskipTests

    to ensure that maven can find the correct classpath and dependencies. Generally, the above command is just a good thing to try running first, if maven is acting oddly.

    16.10.6. Adding Metrics

    After adding a new feature a developer might want to add metrics. HBase exposes metrics using the Hadoop Metrics 2 system, so adding a new metric involves exposing that metric to the hadoop system. Unfortunately the API of metrics2 changed from hadoop 1 to hadoop 2. In order to get around this a set of interfaces and implementations have to be loaded at runtime. To get an in-depth look at the reasoning and structure of these classes you can read the blog post located here. To add a metric to an existing MBean follow the short guide below: Add Metric name and Function to Hadoop Compat Interface.

    Inside of the source interface the corresponds to where the metrics are generated (eg MetricsMasterSource for things coming from HMaster) create new static strings for metric name and description. Then add a new method that will be called to add new reading. Add the Implementation to Both Hadoop 1 and Hadoop 2 Compat modules.

    Inside of the implementation of the source (eg. MetricsMasterSourceImpl in the above example) create a new histogram, counter, gauge, or stat in the init method. Then in the method that was added to the interface wire up the parameter passed in to the histogram.

    Now add tests that make sure the data is correctly exported to the metrics 2 system. For this the MetricsAssertHelper is provided.

    16.11. Submitting Patches

    If you are new to submitting patches to open source or new to submitting patches to Apache, I'd suggest you start by reading the On Contributing Patches page from Apache Commons Project. Its a nice overview that applies equally to the Apache HBase Project.

    16.11.1. Create Patch

    See the aforementioned Apache Commons link for how to make patches against a checked out subversion repository. Patch files can also be easily generated from Eclipse, for example by selecting "Team -> Create Patch". Patches can also be created by git diff and svn diff.

    Please submit one patch-file per Jira. For example, if multiple files are changed make sure the selected resource when generating the patch is a directory. Patch files can reflect changes in multiple files.

    Generating patches using git:

    $ git diff --no-prefix  > HBASE_XXXX.patch

    Don't forget the 'no-prefix' option; and generate the diff from the root directory of project

    Make sure you review Section, “Code Formatting” for code style.

    16.11.2. Patch File Naming

    The patch file should have the Apache HBase Jira ticket in the name. For example, if a patch was submitted for, then a patch file called Foo_HBASE_XXXX.patch would be acceptable where XXXX is the Apache HBase Jira number.

    If you generating from a branch, then including the target branch in the filename is advised, e.g., HBASE_XXXX-0.90.patch.

    16.11.3. Unit Tests

    Yes, please. Please try to include unit tests with every code patch (and especially new classes and large changes). Make sure unit tests pass locally before submitting the patch.

    Also, see Section, “Mockito”.

    If you are creating a new unit test class, notice how other unit test classes have classification/sizing annotations at the top and a static method on the end. Be sure to include these in any new unit test files you generate. See Section 16.7, “Tests” for more on how the annotations work.

    16.11.4. Attach Patch to Jira

    The patch should be attached to the associated Jira ticket "More Actions -> Attach Files". Make sure you click the ASF license inclusion, otherwise the patch can't be considered for inclusion.

    Once attached to the ticket, click "Submit Patch" and the status of the ticket will change. Committers will review submitted patches for inclusion into the codebase. Please understand that not every patch may get committed, and that feedback will likely be provided on the patch. Fear not, though, because the Apache HBase community is helpful!

    16.11.5. Common Patch Feedback

    The following items are representative of common patch feedback. Your patch process will go faster if these are taken into account before submission.

    See the Java coding standards for more information on coding conventions in Java. Space Invaders

    Rather than do this...

    if ( foo.equals( bar ) ) {     // don't do this

    ... do this instead...

    if (foo.equals(bar)) {

    Also, rather than do this...

    foo = barArray[ i ];     // don't do this

    ... do this instead...

    foo = barArray[i];
 Auto Generated Code

    Auto-generated code in Eclipse often looks like this...

     public void readFields(DataInput arg0) throws IOException {    // don't do this
       foo = arg0.readUTF();                                       // don't do this

    ... do this instead ...

     public void readFields(DataInput di) throws IOException {
       foo = di.readUTF();

    See the difference? 'arg0' is what Eclipse uses for arguments by default. Long Lines

    Keep lines less than 100 characters.

    Bar bar = foo.veryLongMethodWithManyArguments(argument1, argument2, argument3, argument4, argument5, argument6, argument7, argument8, argument9);  // don't do this

    ... do something like this instead ...

    Bar bar = foo.veryLongMethodWithManyArguments(
     argument1, argument2, argument3,argument4, argument5, argument6, argument7, argument8, argument9);
 Trailing Spaces

    This happens more than people would imagine.

    Bar bar = foo.getBar();     <--- imagine there's an extra space(s) after the semicolon instead of a line break.

    Make sure there's a line-break after the end of your code, and also avoid lines that have nothing but whitespace. Implementing Writable

    Applies pre-0.96 only

    In 0.96, HBase moved to protobufs. The below section on Writables applies to 0.94.x and previous, not to 0.96 and beyond.

    Every class returned by RegionServers must implement Writable. If you are creating a new class that needs to implement this interface, don't forget the default constructor. Javadoc

    This is also a very common feedback item. Don't forget Javadoc!

    Javadoc warnings are checked during precommit. If the precommit tool gives you a '-1', please fix the javadoc issue. Your patch won't be committed if it adds such warnings. Findbugs

    Findbugs is used to detect common bugs pattern. As Javadoc, it is checked during the precommit build up on Apache's Jenkins, and as with Javadoc, please fix them. You can run findbugs locally with 'mvn findbugs:findbugs': it will generate the findbugs files locally. Sometimes, you may have to write code smarter than Findbugs. You can annotate your code to tell Findbugs you know what you're doing, by annotating your class with:

                        justification="I know what I'm doing")

    Note that we're using the apache licensed version of the annotations. Javadoc - Useless Defaults

    Don't just leave the @param arguments the way your IDE generated them. Don't do this...

       * @param bar             <---- don't do this!!!!
       * @return                <---- or this!!!!
      public Foo getFoo(Bar bar);

    ... either add something descriptive to the @param and @return lines, or just remove them. But the preference is to add something descriptive and useful. One Thing At A Time, Folks

    If you submit a patch for one thing, don't do auto-reformatting or unrelated reformatting of code on a completely different area of code.

    Likewise, don't add unrelated cleanup or refactorings outside the scope of your Jira. Ambigious Unit Tests

    Make sure that you're clear about what you are testing in your unit tests and why.

    16.11.6. Submitting a patch again

    Sometimes committers ask for changes for a patch. After incorporating the suggested/requested changes, follow the following process to submit the patch again.

    • Do not delete the old patch file
    • version your new patch file using a simple scheme like this:
      HBASE-{jira number}-{version}.patch
      e.g: HBASE_XXXX-v2.patch
    • 'Cancel Patch' on JIRA.. bug status will change back to Open
    • Attach new patch file (e.g. HBASE_XXXX-v2.patch) using 'Files --> Attach'
    • Click on 'Submit Patch'. Now the bug status will say 'Patch Available'.
    Committers will review the patch. Rinse and repeat as many times as needed :-)

    16.11.7. Submitting incremental patches

    At times you may want to break a big change into mulitple patches. Here is a sample work-flow using git

    • patch 1:
      • $ git diff --no-prefix > HBASE_XXXX-1.patch
    • patch 2:
      • create a new git branch
        $ git checkout -b my_branch
      • save your work $ git add file1 file2
        $ git commit -am 'saved after HBASE_XXXX-1.patch'
        now you have your own branch, that is different from remote master branch
      • make more changes...
      • create second patch
        $ git diff --no-prefix > HBASE_XXXX-2.patch

    16.11.8. ReviewBoard

    Larger patches should go through ReviewBoard.

    For more information on how to use ReviewBoard, see the ReviewBoard documentation.

    16.11.9. Committing Patches

    Committers do this. See How To Commit in the Apache HBase wiki.

    Commiters will also resolve the Jira, typically after the patch passes a build. Committers are responsible for making sure commits do not break the build or tests

    If a committer commits a patch it is their responsibility to make sure it passes the test suite. It is helpful if contributors keep an eye out that their patch does not break the hbase build and/or tests but ultimately, a contributor cannot be expected to be up on the particular vagaries and interconnections that occur in a project like hbase. A committer should.

    [32] Before 0.95.0, site and reference guide were at src/docbkx and src/site respectively

    [33] There are currently a few exceptions that we need to fix around whether a table is enabled or disabled

    Chapter 17. ZooKeeper

    A distributed Apache HBase installation depends on a running ZooKeeper cluster. All participating nodes and clients need to be able to access the running ZooKeeper ensemble. Apache HBase by default manages a ZooKeeper "cluster" for you. It will start and stop the ZooKeeper ensemble as part of the HBase start/stop process. You can also manage the ZooKeeper ensemble independent of HBase and just point HBase at the cluster it should use. To toggle HBase management of ZooKeeper, use the HBASE_MANAGES_ZK variable in conf/ This variable, which defaults to true, tells HBase whether to start/stop the ZooKeeper ensemble servers as part of HBase start/stop.

    When HBase manages the ZooKeeper ensemble, you can specify ZooKeeper configuration using its native zoo.cfg file, or, the easier option is to just specify ZooKeeper options directly in conf/hbase-site.xml. A ZooKeeper configuration option can be set as a property in the HBase hbase-site.xml XML configuration file by prefacing the ZooKeeper option name with For example, the clientPort setting in ZooKeeper can be changed by setting the property. For all default values used by HBase, including ZooKeeper configuration, see Section, “HBase Default Configuration”. Look for the prefix [34]

    You must at least list the ensemble servers in hbase-site.xml using the hbase.zookeeper.quorum property. This property defaults to a single ensemble member at localhost which is not suitable for a fully distributed HBase. (It binds to the local machine only and remote clients will not be able to connect).

    How many ZooKeepers should I run?

    You can run a ZooKeeper ensemble that comprises 1 node only but in production it is recommended that you run a ZooKeeper ensemble of 3, 5 or 7 machines; the more members an ensemble has, the more tolerant the ensemble is of host failures. Also, run an odd number of machines. In ZooKeeper, an even number of peers is supported, but it is normally not used because an even sized ensemble requires, proportionally, more peers to form a quorum than an odd sized ensemble requires. For example, an ensemble with 4 peers requires 3 to form a quorum, while an ensemble with 5 also requires 3 to form a quorum. Thus, an ensemble of 5 allows 2 peers to fail, and thus is more fault tolerant than the ensemble of 4, which allows only 1 down peer.

    Give each ZooKeeper server around 1GB of RAM, and if possible, its own dedicated disk (A dedicated disk is the best thing you can do to ensure a performant ZooKeeper ensemble). For very heavily loaded clusters, run ZooKeeper servers on separate machines from RegionServers (DataNodes and TaskTrackers).

    For example, to have HBase manage a ZooKeeper quorum on nodes rs{1,2,3,4,5}, bound to port 2222 (the default is 2181) ensure HBASE_MANAGE_ZK is commented out or set to true in conf/ and then edit conf/hbase-site.xml and set and hbase.zookeeper.quorum. You should also set to other than the default as the default has ZooKeeper persist data under /tmp which is often cleared on system restart. In the example below we have ZooKeeper persist to /user/local/zookeeper.

          <description>Property from ZooKeeper's config zoo.cfg.
          The port at which the clients will connect.
          <description>Comma separated list of servers in the ZooKeeper Quorum.
          For example, ",,".
          By default this is set to localhost for local and pseudo-distributed modes
          of operation. For a fully-distributed setup, this should be set to a full
          list of ZooKeeper quorum servers. If HBASE_MANAGES_ZK is set in
          this is the list of servers which we will start/stop ZooKeeper on.
          <description>Property from ZooKeeper's config zoo.cfg.
          The directory where the snapshot is stored.

    What verion of ZooKeeper should I use?

    The newer version, the better. For example, some folks have been bitten by ZOOKEEPER-1277. If running zookeeper 3.5+, you can ask hbase to make use of the new multi operation by enabling hbase.zookeeper.useMulti" in your hbase-site.xml.

    ZooKeeper Maintenance

    Be sure to set up the data dir cleaner described under Zookeeper Maintenance else you could have 'interesting' problems a couple of months in; i.e. zookeeper could start dropping sessions if it has to run through a directory of hundreds of thousands of logs which is wont to do around leader reelection time -- a process rare but run on occasion whether because a machine is dropped or happens to hiccup.

    17.1. Using existing ZooKeeper ensemble

    To point HBase at an existing ZooKeeper cluster, one that is not managed by HBase, set HBASE_MANAGES_ZK in conf/ to false

      # Tell HBase whether it should manage its own instance of Zookeeper or not.
      export HBASE_MANAGES_ZK=false

    Next set ensemble locations and client port, if non-standard, in hbase-site.xml, or add a suitably configured zoo.cfg to HBase's CLASSPATH. HBase will prefer the configuration found in zoo.cfg over any settings in hbase-site.xml.

    When HBase manages ZooKeeper, it will start/stop the ZooKeeper servers as a part of the regular start/stop scripts. If you would like to run ZooKeeper yourself, independent of HBase start/stop, you would do the following

    ${HBASE_HOME}/bin/ {start,stop} zookeeper

    Note that you can use HBase in this manner to spin up a ZooKeeper cluster, unrelated to HBase. Just make sure to set HBASE_MANAGES_ZK to false if you want it to stay up across HBase restarts so that when HBase shuts down, it doesn't take ZooKeeper down with it.

    For more information about running a distinct ZooKeeper cluster, see the ZooKeeper Getting Started Guide. Additionally, see the ZooKeeper Wiki or the ZooKeeper documentation for more information on ZooKeeper sizing.

    17.2. SASL Authentication with ZooKeeper

    Newer releases of Apache HBase (>= 0.92) will support connecting to a ZooKeeper Quorum that supports SASL authentication (which is available in Zookeeper versions 3.4.0 or later).

    This describes how to set up HBase to mutually authenticate with a ZooKeeper Quorum. ZooKeeper/HBase mutual authentication (HBASE-2418) is required as part of a complete secure HBase configuration (HBASE-3025). For simplicity of explication, this section ignores additional configuration required (Secure HDFS and Coprocessor configuration). It's recommended to begin with an HBase-managed Zookeeper configuration (as opposed to a standalone Zookeeper quorum) for ease of learning.

    17.2.1. Operating System Prerequisites

    You need to have a working Kerberos KDC setup. For each $HOST that will run a ZooKeeper server, you should have a principle zookeeper/$HOST. For each such host, add a service key (using the kadmin or kadmin.local tool's ktadd command) for zookeeper/$HOST and copy this file to $HOST, and make it readable only to the user that will run zookeeper on $HOST. Note the location of this file, which we will use below as $PATH_TO_ZOOKEEPER_KEYTAB.

    Similarly, for each $HOST that will run an HBase server (master or regionserver), you should have a principle: hbase/$HOST. For each host, add a keytab file called hbase.keytab containing a service key for hbase/$HOST, copy this file to $HOST, and make it readable only to the user that will run an HBase service on $HOST. Note the location of this file, which we will use below as $PATH_TO_HBASE_KEYTAB.

    Each user who will be an HBase client should also be given a Kerberos principal. This principal should usually have a password assigned to it (as opposed to, as with the HBase servers, a keytab file) which only this user knows. The client's principal's maxrenewlife should be set so that it can be renewed enough so that the user can complete their HBase client processes. For example, if a user runs a long-running HBase client process that takes at most 3 days, we might create this user's principal within kadmin with: addprinc -maxrenewlife 3days. The Zookeeper client and server libraries manage their own ticket refreshment by running threads that wake up periodically to do the refreshment.

    On each host that will run an HBase client (e.g. hbase shell), add the following file to the HBase home directory's conf directory:

                      Client {

    We'll refer to this JAAS configuration file as $CLIENT_CONF below.

    17.2.2. HBase-managed Zookeeper Configuration

    On each node that will run a zookeeper, a master, or a regionserver, create a JAAS configuration file in the conf directory of the node's HBASE_HOME directory that looks like the following:

                      Server {
                      Client {
    where the $PATH_TO_HBASE_KEYTAB and $PATH_TO_ZOOKEEPER_KEYTAB files are what you created above, and $HOST is the hostname for that node.

    The Server section will be used by the Zookeeper quorum server, while the Client section will be used by the HBase master and regionservers. The path to this file should be substituted for the text $HBASE_SERVER_CONF in the listing below.

    The path to this file should be substituted for the text $CLIENT_CONF in the listing below.

    Modify your to include the following:

                      export HBASE_OPTS="$CLIENT_CONF"
                      export HBASE_MANAGES_ZK=true
                      export HBASE_ZOOKEEPER_OPTS="$HBASE_SERVER_CONF"
                      export HBASE_MASTER_OPTS="$HBASE_SERVER_CONF"
                      export HBASE_REGIONSERVER_OPTS="$HBASE_SERVER_CONF"
    where $HBASE_SERVER_CONF and $CLIENT_CONF are the full paths to the JAAS configuration files created above.

    Modify your hbase-site.xml on each node that will run zookeeper, master or regionserver to contain:


    where $ZK_NODES is the comma-separated list of hostnames of the Zookeeper Quorum hosts.

    Start your hbase cluster by running one or more of the following set of commands on the appropriate hosts:

                      bin/hbase zookeeper start
                      bin/hbase master start
                      bin/hbase regionserver start

    17.2.3. External Zookeeper Configuration

    Add a JAAS configuration file that looks like:

                      Client {

    where the $PATH_TO_HBASE_KEYTAB is the keytab created above for HBase services to run on this host, and $HOST is the hostname for that node. Put this in the HBase home's configuration directory. We'll refer to this file's full pathname as $HBASE_SERVER_CONF below.

    Modify your to include the following:

                      export HBASE_OPTS="$CLIENT_CONF"
                      export HBASE_MANAGES_ZK=false
                      export HBASE_MASTER_OPTS="$HBASE_SERVER_CONF"
                      export HBASE_REGIONSERVER_OPTS="$HBASE_SERVER_CONF"

    Modify your hbase-site.xml on each node that will run a master or regionserver to contain:


    where $ZK_NODES is the comma-separated list of hostnames of the Zookeeper Quorum hosts.

    Add a zoo.cfg for each Zookeeper Quorum host containing:


    Also on each of these hosts, create a JAAS configuration file containing:

                      Server {

    where $HOST is the hostname of each Quorum host. We will refer to the full pathname of this file as $ZK_SERVER_CONF below.

    Start your Zookeepers on each Zookeeper Quorum host with:

                        SERVER_JVMFLAGS="$ZK_SERVER_CONF" bin/zkServer start

    Start your HBase cluster by running one or more of the following set of commands on the appropriate nodes:

                      bin/hbase master start
                      bin/hbase regionserver start

    17.2.4. Zookeeper Server Authentication Log Output

    If the configuration above is successful, you should see something similar to the following in your Zookeeper server logs:

    11/12/05 22:43:39 INFO zookeeper.Login: successfully logged in.
    11/12/05 22:43:39 INFO server.NIOServerCnxnFactory: binding to port
    11/12/05 22:43:39 INFO zookeeper.Login: TGT refresh thread started.
    11/12/05 22:43:39 INFO zookeeper.Login: TGT valid starting at:        Mon Dec 05 22:43:39 UTC 2011
    11/12/05 22:43:39 INFO zookeeper.Login: TGT expires:                  Tue Dec 06 22:43:39 UTC 2011
    11/12/05 22:43:39 INFO zookeeper.Login: TGT refresh sleeping until: Tue Dec 06 18:36:42 UTC 2011
    11/12/05 22:43:59 INFO auth.SaslServerCallbackHandler:
      Successfully authenticated client: authenticationID=hbase/;
    11/12/05 22:43:59 INFO auth.SaslServerCallbackHandler: Setting authorizedID: hbase
    11/12/05 22:43:59 INFO server.ZooKeeperServer: adding SASL authorization for authorizationID: hbase

    17.2.5. Zookeeper Client Authentication Log Output

    On the Zookeeper client side (HBase master or regionserver), you should see something similar to the following:

    11/12/05 22:43:59 INFO zookeeper.ZooKeeper: Initiating client connection, sessionTimeout=180000 watcher=master:60000
    11/12/05 22:43:59 INFO zookeeper.ClientCnxn: Opening socket connection to server /
    11/12/05 22:43:59 INFO zookeeper.RecoverableZooKeeper: The identifier of this process is 14851@ip-10-166-175-249
    11/12/05 22:43:59 INFO zookeeper.Login: successfully logged in.
    11/12/05 22:43:59 INFO client.ZooKeeperSaslClient: Client will use GSSAPI as SASL mechanism.
    11/12/05 22:43:59 INFO zookeeper.Login: TGT refresh thread started.
    11/12/05 22:43:59 INFO zookeeper.ClientCnxn: Socket connection established to, initiating session
    11/12/05 22:43:59 INFO zookeeper.Login: TGT valid starting at:        Mon Dec 05 22:43:59 UTC 2011
    11/12/05 22:43:59 INFO zookeeper.Login: TGT expires:                  Tue Dec 06 22:43:59 UTC 2011
    11/12/05 22:43:59 INFO zookeeper.Login: TGT refresh sleeping until: Tue Dec 06 18:30:37 UTC 2011
    11/12/05 22:43:59 INFO zookeeper.ClientCnxn: Session establishment complete on server, sessionid = 0x134106594320000, negotiated timeout = 180000

    17.2.6. Configuration from Scratch

    This has been tested on the current standard Amazon Linux AMI. First setup KDC and principals as described above. Next checkout code and run a sanity check.
                    git clone git://
                    cd hbase
                    mvn clean test -Dtest=TestZooKeeperACL
    Then configure HBase as described above. Manually edit target/cached_classpath.txt (see below)..
                    bin/hbase zookeeper &
                    bin/hbase master &
                    bin/hbase regionserver &

    17.2.7. Future improvements Fix target/cached_classpath.txt

    You must override the standard hadoop-core jar file from the target/cached_classpath.txt file with the version containing the HADOOP-7070 fix. You can use the following script to do this:

                      echo `find ~/.m2 -name "*hadoop-core*7070*SNAPSHOT.jar"` ':' `cat target/cached_classpath.txt` | sed 's/ //g' > target/tmp.txt
                      mv target/tmp.txt target/cached_classpath.txt
                 Set JAAS configuration programmatically

    This would avoid the need for a separate Hadoop jar that fixes HADOOP-7070. Elimination of kerberos.removeHostFromPrincipal and kerberos.removeRealmFromPrincipal

    [34] For the full list of ZooKeeper configurations, see ZooKeeper's zoo.cfg. HBase does not ship with a zoo.cfg so you will need to browse the conf directory in an appropriate ZooKeeper download.

    Chapter 18. Apache HBase Coprocessors

    The idea of HBase coprocessors was inspired by Google's BigTable coprocessors. The Apache HBase Blog on Coprocessor is a very good documentation on that. It has detailed information about the coprocessor framework, terminology, management, and so on.

    Chapter 19. Community

    19.1. Decisions

    19.1.1. Feature Branches

    Feature Branches are easy to make. You do not have to be a committer to make one. Just request the name of your branch be added to JIRA up on the developer's mailing list and a committer will add it for you. Thereafter you can file issues against your feature branch in Apache HBase JIRA. Your code you keep elsewhere -- it should be public so it can be observed -- and you can update dev mailing list on progress. When the feature is ready for commit, 3 +1s from committers will get your feature merged[35]

    19.1.2. Patch +1 Policy

    The below policy is something we put in place 09/2012. It is a suggested policy rather than a hard requirement. We want to try it first to see if it works before we cast it in stone.

    Apache HBase is made of components. Components have one or more Section 19.2.1, “Component Owner/Lieutenant”s. See the 'Description' field on the components JIRA page for who the current owners are by component.

    Patches that fit within the scope of a single Apache HBase component require, at least, a +1 by one of the component's owners before commit. If owners are absent -- busy or otherwise -- two +1s by non-owners will suffice.

    Patches that span components need at least two +1s before they can be committed, preferably +1s by owners of components touched by the x-component patch (TODO: This needs tightening up but I think fine for first pass).

    Any -1 on a patch by anyone vetos a patch; it cannot be committed until the justification for the -1 is addressed.

    19.1.3. How to set fix version in JIRA on issue resolve

    Here is how we agreed to set versions in JIRA when we resolve an issue. If trunk is going to be 0.98.0 then:

    • Commit only to trunk: Mark with 0.98

    • Commit to 0.95 and trunk : Mark with 0.98, and 0.95.x

    • Commit to 0.94.x and 0.95, and trunk: Mark with 0.98, 0.95.x, and 0.94.x

    • Commit to 89-fb: Mark with 89-fb.

    • Commit site fixes: no version

    19.1.4. Policy on when to set a RESOLVED JIRA as CLOSED

    We agreed that for issues that list multiple releases in their Fix Version/s field, CLOSE the issue on the release of any of the versions listed; subsequent change to the issue must happen in a new JIRA.

    19.2. Community Roles

    19.2.1. Component Owner/Lieutenant

    Component owners are listed in the description field on this Apache HBase JIRA components page. The owners are listed in the 'Description' field rather than in the 'Component Lead' field because the latter only allows us list one individual whereas it is encouraged that components have multiple owners.

    Owners or component lieutenants are volunteers who are (usually, but not necessarily) expert in their component domain and may have an agenda on how they think their Apache HBase component should evolve.

    Duties include:

    1. Owners will try and review patches that land within their component's scope.

    2. If applicable, if an owner has an agenda, they will publish their goals or the design toward which they are driving their component

    If you would like to be volunteer as a component owner, just write the dev list and we'll sign you up. Owners do not need to be committers.

    Appendix A. FAQ

    A.1. General
    When should I use HBase?
    Are there other HBase FAQs?
    Does HBase support SQL?
    How can I find examples of NoSQL/HBase?
    What is the history of HBase?
    A.2. Architecture
    How does HBase handle Region-RegionServer assignment and locality?
    A.3. Configuration
    How can I get started with my first cluster?
    Where can I learn about the rest of the configuration options?
    A.4. Schema Design / Data Access
    How should I design my schema in HBase?
    How can I store (fill in the blank) in HBase?
    How can I handle secondary indexes in HBase?
    Can I change a table's rowkeys?
    What APIs does HBase support?
    A.5. MapReduce
    How can I use MapReduce with HBase?
    A.6. Performance and Troubleshooting
    How can I improve HBase cluster performance?
    How can I troubleshoot my HBase cluster?
    A.7. Amazon EC2
    I am running HBase on Amazon EC2 and...
    A.8. Operations
    How do I manage my HBase cluster?
    How do I back up my HBase cluster?
    A.9. HBase in Action
    Where can I find interesting videos and presentations on HBase?

    A.1. General

    When should I use HBase?
    Are there other HBase FAQs?
    Does HBase support SQL?
    How can I find examples of NoSQL/HBase?
    What is the history of HBase?

    When should I use HBase?

    See the Section 9.1, “Overview” in the Architecture chapter.

    Are there other HBase FAQs?

    See the FAQ that is up on the wiki, HBase Wiki FAQ.

    Does HBase support SQL?

    Not really. SQL-ish support for HBase via Hive is in development, however Hive is based on MapReduce which is not generally suitable for low-latency requests. See the Chapter 5, Data Model section for examples on the HBase client.

    How can I find examples of NoSQL/HBase?

    See the link to the BigTable paper in Appendix F, Other Information About HBase in the appendix, as well as the other papers.

    What is the history of HBase?

    See Appendix G, HBase History.

    A.2. Architecture

    How does HBase handle Region-RegionServer assignment and locality?

    How does HBase handle Region-RegionServer assignment and locality?

    See Section 9.7, “Regions”.

    A.3. Configuration

    How can I get started with my first cluster?
    Where can I learn about the rest of the configuration options?

    How can I get started with my first cluster?

    See Section 1.2, “Quick Start”.

    Where can I learn about the rest of the configuration options?

    See Chapter 2, Apache HBase Configuration.

    A.4. Schema Design / Data Access

    How should I design my schema in HBase?
    How can I store (fill in the blank) in HBase?
    How can I handle secondary indexes in HBase?
    Can I change a table's rowkeys?
    What APIs does HBase support?

    How should I design my schema in HBase?

    See Chapter 5, Data Model and Chapter 6, HBase and Schema Design

    How can I store (fill in the blank) in HBase?

    See Section 6.5, “ Supported Datatypes ”.

    How can I handle secondary indexes in HBase?

    See Section 6.9, “ Secondary Indexes and Alternate Query Paths ”

    Can I change a table's rowkeys?

    This is a very common quesiton. You can't. See Section 6.3.5, “Immutability of Rowkeys”.

    What APIs does HBase support?

    See Chapter 5, Data Model, Section 9.3, “Client” and Section 10.1, “Non-Java Languages Talking to the JVM”.

    A.5. MapReduce

    How can I use MapReduce with HBase?

    How can I use MapReduce with HBase?

    See Chapter 7, HBase and MapReduce

    A.6. Performance and Troubleshooting

    How can I improve HBase cluster performance?
    How can I troubleshoot my HBase cluster?

    How can I improve HBase cluster performance?

    See Chapter 12, Apache HBase Performance Tuning.

    How can I troubleshoot my HBase cluster?

    See Chapter 13, Troubleshooting and Debugging Apache HBase.

    A.7. Amazon EC2

    I am running HBase on Amazon EC2 and...

    I am running HBase on Amazon EC2 and...

    EC2 issues are a special case. See Troubleshooting Section 13.12, “Amazon EC2” and Performance Section 12.12, “Amazon EC2” sections.

    A.8. Operations

    How do I manage my HBase cluster?
    How do I back up my HBase cluster?

    How do I manage my HBase cluster?

    See Chapter 15, Apache HBase Operational Management

    How do I back up my HBase cluster?

    See Section 15.7, “HBase Backup”

    A.9. HBase in Action

    Where can I find interesting videos and presentations on HBase?

    Where can I find interesting videos and presentations on HBase?

    See Appendix F, Other Information About HBase

    Appendix B. hbck In Depth

    HBaseFsck (hbck) is a tool for checking for region consistency and table integrity problems and repairing a corrupted HBase. It works in two basic modes -- a read-only inconsistency identifying mode and a multi-phase read-write repair mode.

    B.1. Running hbck to identify inconsistencies

    To check to see if your HBase cluster has corruptions, run hbck against your HBase cluster:
    $ ./bin/hbase hbck

    At the end of the commands output it prints OK or tells you the number of INCONSISTENCIES present. You may also want to run run hbck a few times because some inconsistencies can be transient (e.g. cluster is starting up or a region is splitting). Operationally you may want to run hbck regularly and setup alert (e.g. via nagios) if it repeatedly reports inconsistencies . A run of hbck will report a list of inconsistencies along with a brief description of the regions and tables affected. The using the -details option will report more details including a representative listing of all the splits present in all the tables.

    $ ./bin/hbase hbck -details
    If you just want to know if some tables are corrupted, you can limit hbck to identify inconsistencies in only specific tables. For example the following command would only attempt to check table TableFoo and TableBar. The benefit is that hbck will run in less time.
    $ ./bin/hbase hbck TableFoo TableBar

    B.2. Inconsistencies

    If after several runs, inconsistencies continue to be reported, you may have encountered a corruption. These should be rare, but in the event they occur newer versions of HBase include the hbck tool enabled with automatic repair options.

    There are two invariants that when violated create inconsistencies in HBase:

    • HBase’s region consistency invariant is satisfied if every region is assigned and deployed on exactly one region server, and all places where this state kept is in accordance.
    • HBase’s table integrity invariant is satisfied if for each table, every possible row key resolves to exactly one region.

    Repairs generally work in three phases -- a read-only information gathering phase that identifies inconsistencies, a table integrity repair phase that restores the table integrity invariant, and then finally a region consistency repair phase that restores the region consistency invariant. Starting from version 0.90.0, hbck could detect region consistency problems report on a subset of possible table integrity problems. It also included the ability to automatically fix the most common inconsistency, region assignment and deployment consistency problems. This repair could be done by using the -fix command line option. These problems close regions if they are open on the wrong server or on multiple region servers and also assigns regions to region servers if they are not open.

    Starting from HBase versions 0.90.7, 0.92.2 and 0.94.0, several new command line options are introduced to aid repairing a corrupted HBase. This hbck sometimes goes by the nickname “uberhbck”. Each particular version of uber hbck is compatible with the HBase’s of the same major version (0.90.7 uberhbck can repair a 0.90.4). However, versions <=0.90.6 and versions <=0.92.1 may require restarting the master or failing over to a backup master.

    B.3. Localized repairs

    When repairing a corrupted HBase, it is best to repair the lowest risk inconsistencies first. These are generally region consistency repairs -- localized single region repairs, that only modify in-memory data, ephemeral zookeeper data, or patch holes in the META table. Region consistency requires that the HBase instance has the state of the region’s data in HDFS (.regioninfo files), the region’s row in the .META. table., and region’s deployment/assignments on region servers and the master in accordance. Options for repairing region consistency include:

    • -fixAssignments (equivalent to the 0.90 -fix option) repairs unassigned, incorrectly assigned or multiply assigned regions.
    • -fixMeta which removes meta rows when corresponding regions are not present in HDFS and adds new meta rows if they regions are present in HDFS while not in META.

    To fix deployment and assignment problems you can run this command:

    $ ./bin/hbase hbck -fixAssignments
    To fix deployment and assignment problems as well as repairing incorrect meta rows you can run this command:.
    $ ./bin/hbase hbck -fixAssignments -fixMeta
    There are a few classes of table integrity problems that are low risk repairs. The first two are degenerate (startkey == endkey) regions and backwards regions (startkey > endkey). These are automatically handled by sidelining the data to a temporary directory (/hbck/xxxx). The third low-risk class is hdfs region holes. This can be repaired by using the:
    • -fixHdfsHoles option for fabricating new empty regions on the file system. If holes are detected you can use -fixHdfsHoles and should include -fixMeta and -fixAssignments to make the new region consistent.
    $ ./bin/hbase hbck -fixAssignments -fixMeta -fixHdfsHoles
    Since this is a common operation, we’ve added a the -repairHoles flag that is equivalent to the previous command:
    $ ./bin/hbase hbck -repairHoles
    If inconsistencies still remain after these steps, you most likely have table integrity problems related to orphaned or overlapping regions.

    B.4. Region Overlap Repairs

    Table integrity problems can require repairs that deal with overlaps. This is a riskier operation because it requires modifications to the file system, requires some decision making, and may require some manual steps. For these repairs it is best to analyze the output of a hbck -details run so that you isolate repairs attempts only upon problems the checks identify. Because this is riskier, there are safeguard that should be used to limit the scope of the repairs. WARNING: This is a relatively new and have only been tested on online but idle HBase instances (no reads/writes). Use at your own risk in an active production environment! The options for repairing table integrity violations include:
    • -fixHdfsOrphans option for “adopting” a region directory that is missing a region metadata file (the .regioninfo file).
    • -fixHdfsOverlaps ability for fixing overlapping regions
    When repairing overlapping regions, a region’s data can be modified on the file system in two ways: 1) by merging regions into a larger region or 2) by sidelining regions by moving data to “sideline” directory where data could be restored later. Merging a large number of regions is technically correct but could result in an extremely large region that requires series of costly compactions and splitting operations. In these cases, it is probably better to sideline the regions that overlap with the most other regions (likely the largest ranges) so that merges can happen on a more reasonable scale. Since these sidelined regions are already laid out in HBase’s native directory and HFile format, they can be restored by using HBase’s bulk load mechanism. The default safeguard thresholds are conservative. These options let you override the default thresholds and to enable the large region sidelining feature.
    • -maxMerge <n> maximum number of overlapping regions to merge
    • -sidelineBigOverlaps if more than maxMerge regions are overlapping, sideline attempt to sideline the regions overlapping with the most other regions.
    • -maxOverlapsToSideline <n> if sidelining large overlapping regions, sideline at most n regions.
    Since often times you would just want to get the tables repaired, you can use this option to turn on all repair options:
    • -repair includes all the region consistency options and only the hole repairing table integrity options.
    Finally, there are safeguards to limit repairs to only specific tables. For example the following command would only attempt to check and repair table TableFoo and TableBar.
    $ ./bin/hbase hbck -repair TableFoo TableBar

    B.4.1. Special cases: Meta is not properly assigned

    There are a few special cases that hbck can handle as well. Sometimes the meta table’s only region is inconsistently assigned or deployed. In this case there is a special -fixMetaOnly option that can try to fix meta assignments.
    $ ./bin/hbase hbck -fixMetaOnly -fixAssignments

    B.4.2. Special cases: HBase version file is missing

    HBase’s data on the file system requires a version file in order to start. If this flie is missing, you can use the -fixVersionFile option to fabricating a new HBase version file. This assumes that the version of hbck you are running is the appropriate version for the HBase cluster.

    B.4.3. Special case: Root and META are corrupt.

    The most drastic corruption scenario is the case where the ROOT or META is corrupted and HBase will not start. In this case you can use the OfflineMetaRepair tool create new ROOT and META regions and tables. This tool assumes that HBase is offline. It then marches through the existing HBase home directory, loads as much information from region metadata files (.regioninfo files) as possible from the file system. If the region metadata has proper table integrity, it sidelines the original root and meta table directories, and builds new ones with pointers to the region directories and their data.
    $ ./bin/hbase org.apache.hadoop.hbase.util.hbck.OfflineMetaRepair
    NOTE: This tool is not as clever as uberhbck but can be used to bootstrap repairs that uberhbck can complete. If the tool succeeds you should be able to start hbase and run online repairs if necessary.

    B.4.4. Special cases: Offline split parent

    Once a region is split, the offline parent will be cleaned up automatically. Sometimes, daughter regions are split again before their parents are cleaned up. HBase can clean up parents in the right order. However, there could be some lingering offline split parents sometimes. They are in META, in HDFS, and not deployed. But HBase can't clean them up. In this case, you can use the -fixSplitParents option to reset them in META to be online and not split. Therefore, hbck can merge them with other regions if fixing overlapping regions option is used.

    This option should not normally be used, and it is not in -fixAll.

    Appendix C. Compression In HBase

    There are a bunch of compression options in HBase. There is some helpful discussion to be found in Documenting Guidance on compression and codecs.

    C.1. CompressionTest Tool

    HBase includes a tool to test compression is set up properly. To run it, type /bin/hbase org.apache.hadoop.hbase.util.CompressionTest. This will emit usage on how to run the tool.

    You need to restart regionserver for it to pick up fixed codecs!

    Be aware that the regionserver caches the result of the compression check it runs ahead of each region open. This means that you will have to restart the regionserver for it to notice that you have fixed any codec issues.

    C.2.  hbase.regionserver.codecs

    To have a RegionServer test a set of codecs and fail-to-start if any code is missing or misinstalled, add the configuration hbase.regionserver.codecs to your hbase-site.xml with a value of codecs to test on startup. For example if the hbase.regionserver.codecs value is lzo,gz and if lzo is not present or improperly installed, the misconfigured RegionServer will fail to start.

    Administrators might make use of this facility to guard against the case where a new server is added to cluster but the cluster requires install of a particular coded.

    C.3.  LZO

    Unfortunately, HBase cannot ship with LZO because of the licensing issues; HBase is Apache-licensed, LZO is GPL. Therefore LZO install is to be done post-HBase install. See the Using LZO Compression wiki page for how to make LZO work with HBase.

    A common problem users run into when using LZO is that while initial setup of the cluster runs smooth, a month goes by and some sysadmin goes to add a machine to the cluster only they'll have forgotten to do the LZO fixup on the new machine. In versions since HBase 0.90.0, we should fail in a way that makes it plain what the problem is, but maybe not.

    See Section C.2, “ hbase.regionserver.codecs for a feature to help protect against failed LZO install.

    C.4.  GZIP

    GZIP will generally compress better than LZO though slower. For some setups, better compression may be preferred. Java will use java's GZIP unless the native Hadoop libs are available on the CLASSPATH; in this case it will use native compressors instead (If the native libs are NOT present, you will see lots of Got brand-new compressor reports in your logs; see ???).

    C.5.  SNAPPY

    If snappy is installed, HBase can make use of it (courtesy of hadoop-snappy [36]).

    1. Build and install snappy on all nodes of your cluster (see below)

    2. Use CompressionTest to verify snappy support is enabled and the libs can be loaded ON ALL NODES of your cluster:

      $ hbase org.apache.hadoop.hbase.util.CompressionTest hdfs://host/path/to/hbase snappy

    3. Create a column family with snappy compression and verify it in the hbase shell:

      $ hbase> create 't1', { NAME => 'cf1', COMPRESSION => 'SNAPPY' }
      hbase> describe 't1'

      In the output of the "describe" command, you need to ensure it lists "COMPRESSION => 'SNAPPY'"

    C.5.1.  Installation

    You will find the snappy library file under the .libs directory from your Snappy build (For example /home/hbase/snappy-1.0.5/.libs/). The file is called where 1.x.x is the version of the snappy code you are building. You can either copy this file into your hbase lib directory -- under lib/native/PLATFORM -- naming the file as, or simply create a symbolic link to it (See ./bin/hbase for how it does library path for native libs).

    The second file you need is the hadoop native library. You will find this file in your hadoop installation directory under lib/native/Linux-amd64-64/ or lib/native/Linux-i386-32/. The file you are looking for is Again, you can simply copy this file or link to it from under hbase in lib/native/PLATFORM (e.g. Linux-amd64-64, etc.), using the name

    At the end of the installation, you should have both and links or files present into lib/native/Linux-amd64-64 or into lib/native/Linux-i386-32 (where the last part of the directory path is the PLATFORM you built and rare running the native lib on)

    To point hbase at snappy support, in set

    export HBASE_LIBRARY_PATH=/pathtoyourhadoop/lib/native/Linux-amd64-64

    In /pathtoyourhadoop/lib/native/Linux-amd64-64 you should have something like:


    C.6. Changing Compression Schemes

    A frequent question on the dist-list is how to change compression schemes for ColumnFamilies. This is actually quite simple, and can be done via an alter command. Because the compression scheme is encoded at the block-level in StoreFiles, the table does not need to be re-created and the data does not copied somewhere else. Just make sure the old codec is still available until you are sure that all of the old StoreFiles have been compacted.

    [36] See Alejandro's note up on the list on difference between Snappy in Hadoop and Snappy in HBase

    TODO: Describe how YCSB is poor for putting up a decent cluster load.

    TODO: Describe setup of YCSB for HBase

    Ted Dunning redid YCSB so it's mavenized and added facility for verifying workloads. See Ted Dunning's YCSB.

    Appendix E. HFile format version 2

    E.1. Motivation

    Note: this feature was introduced in HBase 0.92

    We found it necessary to revise the HFile format after encountering high memory usage and slow startup times caused by large Bloom filters and block indexes in the region server. Bloom filters can get as large as 100 MB per HFile, which adds up to 2 GB when aggregated over 20 regions. Block indexes can grow as large as 6 GB in aggregate size over the same set of regions. A region is not considered opened until all of its block index data is loaded. Large Bloom filters produce a different performance problem: the first get request that requires a Bloom filter lookup will incur the latency of loading the entire Bloom filter bit array.

    To speed up region server startup we break Bloom filters and block indexes into multiple blocks and write those blocks out as they fill up, which also reduces the HFile writer’s memory footprint. In the Bloom filter case, “filling up a block” means accumulating enough keys to efficiently utilize a fixed-size bit array, and in the block index case we accumulate an “index block” of the desired size. Bloom filter blocks and index blocks (we call these “inline blocks”) become interspersed with data blocks, and as a side effect we can no longer rely on the difference between block offsets to determine data block length, as it was done in version 1.

    HFile is a low-level file format by design, and it should not deal with application-specific details such as Bloom filters, which are handled at StoreFile level. Therefore, we call Bloom filter blocks in an HFile "inline" blocks. We also supply HFile with an interface to write those inline blocks.

    Another format modification aimed at reducing the region server startup time is to use a contiguous “load-on-open” section that has to be loaded in memory at the time an HFile is being opened. Currently, as an HFile opens, there are separate seek operations to read the trailer, data/meta indexes, and file info. To read the Bloom filter, there are two more seek operations for its “data” and “meta” portions. In version 2, we seek once to read the trailer and seek again to read everything else we need to open the file from a contiguous block.

    E.2. HFile format version 1 overview

    As we will be discussing the changes we are making to the HFile format, it is useful to give a short overview of the previous (HFile version 1) format. An HFile in the existing format is structured as follows: HFile Version 1 [37]

    E.2.1.  Block index format in version 1

    The block index in version 1 is very straightforward. For each entry, it contains:

    1. Offset (long)

    2. Uncompressed size (int)

    3. Key (a serialized byte array written using Bytes.writeByteArray)

      1. Key length as a variable-length integer (VInt)

      2. Key bytes

    The number of entries in the block index is stored in the fixed file trailer, and has to be passed in to the method that reads the block index. One of the limitations of the block index in version 1 is that it does not provide the compressed size of a block, which turns out to be necessary for decompression. Therefore, the HFile reader has to infer this compressed size from the offset difference between blocks. We fix this limitation in version 2, where we store on-disk block size instead of uncompressed size, and get uncompressed size from the block header.

    E.3.  HBase file format with inline blocks (version 2)

    E.3.1.  Overview

    The version of HBase introducing the above features reads both version 1 and 2 HFiles, but only writes version 2 HFiles. A version 2 HFile is structured as follows: HFile Version 2

    E.3.2. Unified version 2 block format

    In the version 2 every block in the data section contains the following fields:

    1. 8 bytes: Block type, a sequence of bytes equivalent to version 1's "magic records". Supported block types are:

      1. DATA – data blocks

      2. LEAF_INDEX – leaf-level index blocks in a multi-level-block-index

      3. BLOOM_CHUNK – Bloom filter chunks

      4. META – meta blocks (not used for Bloom filters in version 2 anymore)

      5. INTERMEDIATE_INDEX – intermediate-level index blocks in a multi-level blockindex

      6. ROOT_INDEX – root>level index blocks in a multi>level block index

      7. FILE_INFO – the “file info” block, a small key>value map of metadata

      8. BLOOM_META – a Bloom filter metadata block in the load>on>open section

      9. TRAILER – a fixed>size file trailer. As opposed to the above, this is not an HFile v2 block but a fixed>size (for each HFile version) data structure

      10. INDEX_V1 – this block type is only used for legacy HFile v1 block

    2. Compressed size of the block's data, not including the header (int).

      Can be used for skipping the current data block when scanning HFile data.

    3. Uncompressed size of the block's data, not including the header (int)

      This is equal to the compressed size if the compression algorithm is NON

    4. File offset of the previous block of the same type (long)

      Can be used for seeking to the previous data/index block

    5. Compressed data (or uncompressed data if the compression algorithm is NONE).

    The above format of blocks is used in the following HFile sections:

    1. Scanned block section. The section is named so because it contains all data blocks that need to be read when an HFile is scanned sequentially.  Also contains leaf block index and Bloom chunk blocks.

    2. Non-scanned block section. This section still contains unified-format v2 blocks but it does not have to be read when doing a sequential scan. This section contains “meta” blocks and intermediate-level index blocks.

    We are supporting “meta” blocks in version 2 the same way they were supported in version 1, even though we do not store Bloom filter data in these blocks anymore.

    E.3.3.  Block index in version 2

    There are three types of block indexes in HFile version 2, stored in two different formats (root and non-root):

    1. Data index — version 2 multi-level block index, consisting of:

      1. Version 2 root index, stored in the data block index section of the file

      2. Optionally, version 2 intermediate levels, stored in the non%root format in the data index section of the file. Intermediate levels can only be present if leaf level blocks are present

      3. Optionally, version 2 leaf levels, stored in the non%root format inline with data blocks

    2. Meta index — version 2 root index format only, stored in the meta index section of the file

    3. Bloom index — version 2 root index format only, stored in the “load-on-open” section as part of Bloom filter metadata.

    E.3.4.  Root block index format in version 2

    This format applies to:

    1. Root level of the version 2 data index

    2. Entire meta and Bloom indexes in version 2, which are always single-level.

    A version 2 root index block is a sequence of entries of the following format, similar to entries of a version 1 block index, but storing on-disk size instead of uncompressed size.

    1. Offset (long)

      This offset may point to a data block or to a deeper>level index block.

    2. On-disk size (int)

    3. Key (a serialized byte array stored using Bytes.writeByteArray)

      1. Key (VInt)

      2. Key bytes

    A single-level version 2 block index consists of just a single root index block. To read a root index block of version 2, one needs to know the number of entries. For the data index and the meta index the number of entries is stored in the trailer, and for the Bloom index it is stored in the compound Bloom filter metadata.

    For a multi-level block index we also store the following fields in the root index block in the load-on-open section of the HFile, in addition to the data structure described above:

    1. Middle leaf index block offset

    2. Middle leaf block on-disk size (meaning the leaf index block containing the reference to the “middle” data block of the file)

    3. The index of the mid-key (defined below) in the middle leaf-level block.

    These additional fields are used to efficiently retrieve the mid-key of the HFile used in HFile splits, which we define as the first key of the block with a zero-based index of (n – 1) / 2, if the total number of blocks in the HFile is n. This definition is consistent with how the mid-key was determined in HFile version 1, and is reasonable in general, because blocks are likely to be the same size on average, but we don’t have any estimates on individual key/value pair sizes.

    When writing a version 2 HFile, the total number of data blocks pointed to by every leaf-level index block is kept track of. When we finish writing and the total number of leaf-level blocks is determined, it is clear which leaf-level block contains the mid-key, and the fields listed above are computed.  When reading the HFile and the mid-key is requested, we retrieve the middle leaf index block (potentially from the block cache) and get the mid-key value from the appropriate position inside that leaf block.

    E.3.5.  Non-root block index format in version 2

    This format applies to intermediate-level and leaf index blocks of a version 2 multi-level data block index. Every non-root index block is structured as follows.

    1. numEntries: the number of entries (int).

    2. entryOffsets: the “secondary index” of offsets of entries in the block, to facilitate a quick binary search on the key (numEntries + 1 int values). The last value is the total length of all entries in this index block. For example, in a non-root index block with entry sizes 60, 80, 50 the “secondary index” will contain the following int array: {0, 60, 140, 190}.

    3. Entries. Each entry contains:

      1. Offset of the block referenced by this entry in the file (long)

      2. On>disk size of the referenced block (int)

      3. Key. The length can be calculated from entryOffsets.

    E.3.6.  Bloom filters in version 2

    In contrast with version 1, in a version 2 HFile Bloom filter metadata is stored in the load-on-open section of the HFile for quick startup.

    1. A compound Bloom filter.

      1. Bloom filter version = 3 (int). There used to be a DynamicByteBloomFilter class that had the Bloom filter version number 2

      2. The total byte size of all compound Bloom filter chunks (long)

      3. Number of hash functions (int

      4. Type of hash functions (int)

      5. The total key count inserted into the Bloom filter (long)

      6. The maximum total number of keys in the Bloom filter (long)

      7. The number of chunks (int)

      8. Comparator class used for Bloom filter keys, a UTF>8 encoded string stored using Bytes.writeByteArray

      9. Bloom block index in the version 2 root block index format

    E.3.7. File Info format in versions 1 and 2

    The file info block is a serialized HbaseMapWritable (essentially a map from byte arrays to byte arrays) with the following keys, among others. StoreFile-level logic adds more keys to this.


    The last key of the file (byte array)


    The average key length in the file (int)


    The average value length in the file (int)

    File info format did not change in version 2. However, we moved the file info to the final section of the file, which can be loaded as one block at the time the HFile is being opened. Also, we do not store comparator in the version 2 file info anymore. Instead, we store it in the fixed file trailer. This is because we need to know the comparator at the time of parsing the load-on-open section of the HFile.

    E.3.8.  Fixed file trailer format differences between versions 1 and 2

    The following table shows common and different fields between fixed file trailers in versions 1 and 2. Note that the size of the trailer is different depending on the version, so it is “fixed” only within one version. However, the version is always stored as the last four-byte integer in the file.

    Version 1

    Version 2

    File info offset (long)

    Data index offset (long)

    loadOnOpenOffset (long)

    The offset of the section that we need toload when opening the file.

    Number of data index entries (int)

    metaIndexOffset (long)

    This field is not being used by the version 1 reader, so we removed it from version 2.

    uncompressedDataIndexSize (long)

    The total uncompressed size of the whole data block index, including root-level, intermediate-level, and leaf-level blocks.

    Number of meta index entries (int)

    Total uncompressed bytes (long)

    numEntries (int)

    numEntries (long)

    Compression codec: 0 = LZO, 1 = GZ, 2 = NONE (int)

    The number of levels in the data block index (int)

    firstDataBlockOffset (long)

    The offset of the first first data block. Used when scanning.

    lastDataBlockEnd (long)

    The offset of the first byte after the last key/value data block. We don't need to go beyond this offset when scanning.

    Version: 1 (int)

    Version: 2 (int)

    E.3.9. getShortMidpointKey(an optimization for data index block)

    Note: this optimization was introduced in HBase 0.95+

    HFiles contain many blocks that contain a range of sorted Cells. Each cell has a key. To save IO when reading Cells, the HFile also has an index that maps a Cell's start key to the offset of the beginning of a particular block. Prior to this optimization, HBase would use the key of the first cell in each data block as the index key.

    In HBASE-7845, we generate a new key that is lexicographically larger than the last key of the previous block and lexicographically equal or smaller than the start key of the current block. While actual keys can potentially be very long, this "fake key" or "virtual key" can be much shorter. For example, if the stop key of previous block is "the quick brown fox", the start key of current block is "the who", we could use "the r" as our virtual key in our hfile index.

    There are two benefits to this:

    • having shorter keys reduces the hfile index size, (allowing us to keep more indexes in memory), and
    • using something closer to the end key of the previous block allows us to avoid a potential extra IO when the target key lives in between the "virtual key" and the key of the first element in the target block.

    This optimization (implemented by the getShortMidpointKey method) is inspired by LevelDB's ByteWiseComparatorImpl::FindShortestSeparator() and FindShortSuccessor().

    [37] Image courtesy of Lars George, hbase-architecture-101-storage.html.

    Appendix F. Other Information About HBase

    F.1. HBase Videos

    Introduction to HBase

    Building Real Time Services at Facebook with HBase by Jonathan Gray (Hadoop World 2011).

    HBase and Hadoop, Mixing Real-Time and Batch Processing at StumbleUpon by JD Cryans (Hadoop World 2010).

    F.2. HBase Presentations (Slides)

    Advanced HBase Schema Design by Lars George (Hadoop World 2011).

    Introduction to HBase by Todd Lipcon (Chicago Data Summit 2011).

    Getting The Most From Your HBase Install by Ryan Rawson, Jonathan Gray (Hadoop World 2009).

    F.3. HBase Papers

    BigTable by Google (2006).

    HBase and HDFS Locality by Lars George (2010).

    No Relation: The Mixed Blessings of Non-Relational Databases by Ian Varley (2009).

    F.4. HBase Sites

    Cloudera's HBase Blog has a lot of links to useful HBase information.

    • CAP Confusion is a relevant entry for background information on distributed storage systems.

    HBase Wiki has a page with a number of presentations.

    HBase RefCard from DZone.

    F.5. HBase Books

    HBase: The Definitive Guide by Lars George.

    F.6. Hadoop Books

    Hadoop: The Definitive Guide by Tom White.

    Appendix G. HBase History

    • 2006: BigTable paper published by Google.
    • 2006 (end of year): HBase development starts.
    • 2008: HBase becomes Hadoop sub-project.
    • 2010: HBase becomes Apache top-level project.

    Appendix H. HBase and the Apache Software Foundation

    HBase is a project in the Apache Software Foundation and as such there are responsibilities to the ASF to ensure a healthy project.

    H.1. ASF Development Process

    See the Apache Development Process page for all sorts of information on how the ASF is structured (e.g., PMC, committers, contributors), to tips on contributing and getting involved, and how open-source works at ASF.

    H.2. ASF Board Reporting

    Once a quarter, each project in the ASF portfolio submits a report to the ASF board. This is done by the HBase project lead and the committers. See ASF board reporting for more information.

    Appendix I. Enabling Dapper-like Tracing in HBase

    HBASE-6449 added support for tracing requests through HBase, using the open source tracing library, HTrace. Setting up tracing is quite simple, however it currently requires some very minor changes to your client code (it would not be very difficult to remove this requirement).

    I.1. SpanReceivers

    The tracing system works by collecting information in structs called ‘Spans’. It is up to you to choose how you want to receive this information by implementing the SpanReceiver interface, which defines one method:

    public void receiveSpan(Span span);

    This method serves as a callback whenever a span is completed. HTrace allows you to use as many SpanReceivers as you want so you can easily send trace information to multiple destinations.

    Configure what SpanReceivers you’d like to use by putting a comma separated list of the fully-qualified class name of classes implementing SpanReceiver in hbase-site.xml property: hbase.trace.spanreceiver.classes.

    HBase includes a HBaseLocalFileSpanReceiver that writes all span information to local files in a JSON-based format. The HBaseLocalFileSpanReceiver looks in hbase-site.xml for a hbase.trace.spanreceiver.localfilespanreceiver.filename property with a value describing the name of the file to which nodes should write their span information.

    If you do not want to use the included HBaseLocalFileSpanReceiver, you are encouraged to write your own receiver (take a look at HBaseLocalFileSpanReceiver for an example). If you think others would benefit from your receiver, file a JIRA or send a pull request to HTrace.

    I.2. Client Modifications

    Currently, you must turn on tracing in your client code. To do this, you simply turn on tracing for requests you think are interesting, and turn it off when the request is done.

    For example, if you wanted to trace all of your get operations, you change this:

    HTable table = new HTable(...);
    Get get = new Get(...);


    Span getSpan = Trace.startSpan(“doing get”, Sampler.ALWAYS);
    try {
      HTable table = new HTable(...);
      Get get = new Get(...);
    } finally {

    If you wanted to trace half of your ‘get’ operations, you would pass in:

    new ProbabilitySampler(0.5)

    in lieu of Sampler.ALWAYS to Trace.startSpan(). See the HTrace README for more information on Samplers.

    Appendix J. 0.95 RPC Specification

    In 0.95, all client/server communication is done with protobuf’ed Messages rather than with Hadoop Writables. Our RPC wire format therefore changes. This document describes the client/server request/response protocol and our new RPC wire-format.

    For what RPC is like in 0.94 and previous, see Benoît/Tsuna’s Unofficial Hadoop / HBase RPC protocol documentation. For more background on how we arrived at this spec., see HBase RPC: WIP

    J.1. Goals

    1. A wire-format we can evolve

    2. A format that does not require our rewriting server core or radically changing its current architecture (for later).

    J.2. TODO

    1. List of problems with currently specified format and where we would like to go in a version2, etc. For example, what would we have to change if anything to move server async or to support streaming/chunking?

    2. Diagram on how it works

    3. A grammar that succinctly describes the wire-format. Currently we have these words and the content of the rpc protobuf idl but a grammar for the back and forth would help with groking rpc. Also, a little state machine on client/server interactions would help with understanding (and ensuring correct implementation).

    J.3. RPC

    The client will send setup information on connection establish. Thereafter, the client invokes methods against the remote server sending a protobuf Message and receiving a protobuf Message in response. Communication is synchronous. All back and forth is preceded by an int that has the total length of the request/response. Optionally, Cells(KeyValues) can be passed outside of protobufs in follow-behind Cell blocks (because we can’t protobuf megabytes of KeyValues or Cells). These CellBlocks are encoded and optionally compressed.

    For more detail on the protobufs involved, see the RPC.proto file in trunk.

    J.3.1. Connection Setup

    Client initiates connection.

    J.3.1.1. Client

    On connection setup, client sends a preamble followed by a connection header.

    J. <preamble>
    <MAGIC 4 byte integer> <1 byte RPC Format Version> <1 byte auth type>[38]

    E.g.: HBas0x000x50 -- 4 bytes of MAGIC -- ‘HBas’ -- plus one-byte of version, 0 in this case, and one byte, 0x50 (SIMPLE). of an auth type.

    J. <Protobuf ConnectionHeader Message>

    Has user info, and “protocol”, as well as the encoders and compression the client will use sending CellBlocks. CellBlock encoders and compressors are for the life of the connection. CellBlock encoders implement org.apache.hadoop.hbase.codec.Codec. CellBlocks may then also be compressed. Compressors implement This protobuf is written using writeDelimited so is prefaced by a pb varint with its serialized length

    J.3.1.2. Server

    After client sends preamble and connection header, server does NOT respond if successful connection setup. No response means server is READY to accept requests and to give out response. If the version or authentication in the preamble is not agreeable or the server has trouble parsing the preamble, it will throw a org.apache.hadoop.hbase.ipc.FatalConnectionException explaining the error and will then disconnect. If the client in the connection header -- i.e. the protobuf’d Message that comes after the connection preamble -- asks for for a Service the server does not support or a codec the server does not have, again we throw a FatalConnectionException with explanation.

    J.3.2. Request

    After a Connection has been set up, client makes requests. Server responds.

    A request is made up of a protobuf RequestHeader followed by a protobuf Message parameter. The header includes the method name and optionally, metadata on the optional CellBlock that may be following. The parameter type suits the method being invoked: i.e. if we are doing a getRegionInfo request, the protobuf Message param will be an instance of GetRegionInfoRequest. The response will be a GetRegionInfoResponse. The CellBlock is optionally used ferrying the bulk of the RPC data: i.e Cells/KeyValues.

    J.3.2.1. Request Parts

    J. <Total Length>

    The request is prefaced by an int that holds the total length of what follows.

    J. <Protobuf RequestHeader Message>

    Will have,, and method name, etc. including optional Metadata on the Cell block IFF one is following. Data is protobuf’d inline in this pb Message or optionally comes in the following CellBlock

    J. <Protobuf Param Message>

    If the method being invoked is getRegionInfo, if you study the Service descriptor for the client to regionserver protocol, you will find that the request sends a GetRegionInfoRequest protobuf Message param in this position.

    J. <CellBlock>

    An encoded and optionally compressed Cell block.

    J.3.3. Response

    Same as Request, it is a protobuf ResponseHeader followed by a protobuf Message response where the Message response type suits the method invoked. Bulk of the data may come in a following CellBlock.

    J.3.3.1. Response Parts

    J. <Total Length>

    The response is prefaced by an int that holds the total length of what follows.

    J. <Protobuf ResponseHeader Message>

    Will have, etc. Will include exception if failed processing.  Optionally includes metadata on optional, IFF there is a CellBlock following.

    J. <Protobuf Response Message>

    Return or may be nothing if exception. If the method being invoked is getRegionInfo, if you study the Service descriptor for the client to regionserver protocol, you will find that the response sends a GetRegionInfoResponse protobuf Message param in this position.

    J. <CellBlock>

    An encoded and optionally compressed Cell block.

    J.3.4. Exceptions

    There are two distinct types. There is the request failed which is encapsulated inside the response header for the response. The connection stays open to receive new requests. The second type, the FatalConnectionException, kills the connection.

    Exceptions can carry extra information. See the ExceptionResponse protobuf type. It has a flag to indicate do-no-retry as well as other miscellaneous payload to help improve client responsiveness.

    J.3.5. CellBlocks

    These are not versioned. Server can do the codec or it cannot. If new version of a codec with say, tighter encoding, then give it a new class name. Codecs will live on the server for all time so old clients can connect.

    J.4. Notes

    J.4.1. Constraints

    In some part, current wire-format -- i.e. all requests and responses preceeded by a length -- has been dictated by current server non-async architecture.

    J.4.2. One fat pb request or header+param

    We went with pb header followed by pb param making a request and a pb header followed by pb response for now. Doing header+param rather than a single protobuf Message with both header and param content:

    1. Is closer to what we currently have

    2. Having a single fat pb requires extra copying putting the already pb’d param into the body of the fat request pb (and same making result)

    3. We can decide whether to accept the request or not before we read the param; for example, the request might be low priority.  As is, we read header+param in one go as server is currently implemented so this is a TODO.

    The advantages are minor.  If later, fat request has clear advantage, can roll out a v2 later.

    J.4.3. RPC Configurations

    J.4.3.1. CellBlock Codecs

    To enable a codec other than the default KeyValueCodec, set hbase.client.rpc.codec to the name of the Codec class to use. Codec must implement hbase's Codec Interface. After connection setup, all passed cellblocks will be sent with this codec. The server will return cellblocks using this same codec as long as the codec is on the servers' CLASSPATH (else you will get UnsupportedCellCodecException).

    To change the default codec, set hbase.client.default.rpc.codec.

    To disable cellblocks completely and to go pure protobuf, set the default to the empty String and do not specify a codec in your Configuration. So, set hbase.client.default.rpc.codec to the empty string and do not set hbase.client.rpc.codec. This will cause the client to connect to the server with no codec specified. If a server sees no codec, it will return all responses in pure protobuf. Running pure protobuf all the time will be slower than running with cellblocks.

    J.4.3.2. Compression

    Uses hadoops compression codecs. To enable compressing of passed CellBlocks, set hbase.client.rpc.compressor to the name of the Compressor to use. Compressor must implement Hadoops' CompressionCodec Interface. After connection setup, all passed cellblocks will be sent compressed. The server will return cellblocks compressed using this same compressor as long as the compressor is on its CLASSPATH (else you will get UnsupportedCompressionCodecException).

    [38] We need the auth method spec. here so the connection header is encoded if auth enabled.



    Cells, Cells
    Column Family, Column Family
    Column Family Qualifier, Column Family
    Compression, Compression In HBase


    Hadoop, Hadoop


    IntegrationTests, Integration Tests


    LargeTests, Large Tests


    MediumTests, Medium Tests
    MSLAB, Long GC pauses


    SmallTests, Small Tests


    Test Resource Checker, Test Resource Checker


    Versions, Versions


    ZooKeeper, ZooKeeper
    comments powered by Disqus