The Apache HBase™ Reference Guide

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-0.94.27.tar.gz
$ cd hbase-0.94.27

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 hbase.zookeeper.property.dataDir, the director ZooKeeper writes its data too:

<?xml version="1.0"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>
<configuration>
  <property>
    <name>hbase.rootdir</name>
    <value>file:///DIRECTORY/hbase</value>
  </property>
  <property>
    <name>hbase.zookeeper.property.dataDir</name>
    <value>/DIRECTORY/zookeeper</value>
  </property>
</configuration>

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-${user.name} 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).

Now start HBase:

$ ./bin/start-hbase.sh
starting Master, logging to logs/hbase-user-master-example.org.out

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.

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

hbase(main):001:0> 

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'
ROW        COLUMN+CELL
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'
COLUMN      CELL
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

Stop your hbase instance by running the stop script.

$ ./bin/stop-hbase.sh
stopping hbase...............

The above described standalone setup is good for testing and experiments only. In the next chapter, Chapter 2, Apache HBase (TM) 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.

Just like Hadoop, HBase requires at least java 6 from Oracle.

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 http://wiki.apache.org/hadoop/Distributions%20and%20Commercial%20Support

Where

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.

  <property>
    <name>dfs.support.append</name>
    <value>true</value>
  </property>
        

      <property>
        <name>dfs.datanode.max.xcievers</name>
        <value>4096</value>
      </property>
      

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.

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 ^[9].

Distributed modes require an instance of the Hadoop Distributed File System (HDFS). See the Hadoop requirements and instructions for how to set up a HDFS. Before proceeding, ensure you have an appropriate, working 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.

<configuration>
  ...
  <property>
    <name>hbase.rootdir</name>
    <value>hdfs://h-24-30.sfo.stumble.net:8020/hbase</value>
  </property>
  <property>
    <name>hbase.cluster.distributed</name>
    <value>true</value>
  </property>
  <property>
    <name>hbase.zookeeper.quorum</name>
    <value>h-24-30.sfo.stumble.net</value>
  </property>
  ...
</configuration>

% bin/start-hbase.sh

% bin/local-master-backup.sh start 1

% bin/local-master-backup.sh start 2 3

% bin/local-regionservers.sh start 1

% bin/local-regionservers.sh start 2 3 4 5

% cat /tmp/hbase-${USER}-1-master.pid |xargs kill -9

% bin/local-regionservers.sh stop 1
	                

<configuration>
  ...
  <property>
    <name>hbase.rootdir</name>
    <value>hdfs://namenode.example.org:8020/hbase</value>
    <description>The directory shared by RegionServers.
    </description>
  </property>
  <property>
    <name>hbase.cluster.distributed</name>
    <value>true</value>
    <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 hbase-env.sh)
    </description>
  </property>
  ...
</configuration>

Make sure HDFS is running first. Start and stop the Hadoop HDFS daemons by running bin/start-hdfs.sh 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:

bin/start-hbase.sh

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 master.example.org on the default port, to see the Master's homepage you'd point your browser at http://master.example.org:60010.

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/stop-hbase.sh
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.

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 ??? 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.

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/hbase-env.sh 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.

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"?>
<configuration>
  <property>
    <name>hbase.zookeeper.quorum</name>
    <value>example1,example2,example3</value>
    <description>The directory shared by region servers.
    </description>
  </property>
</configuration>

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

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 hbase-env.sh -- found in the HBase conf directory might look like.

<?xml version="1.0"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>
<configuration>
  <property>
    <name>hbase.zookeeper.quorum</name>
    <value>example1,example2,example3</value>
    <description>The directory shared by RegionServers.
    </description>
  </property>
  <property>
    <name>hbase.zookeeper.property.dataDir</name>
    <value>/export/zookeeper</value>
    <description>Property from ZooKeeper's config zoo.cfg.
    The directory where the snapshot is stored.
    </description>
  </property>
  <property>
    <name>hbase.rootdir</name>
    <value>hdfs://example0:8020/hbase</value>
    <description>The directory shared by RegionServers.
    </description>
  </property>
  <property>
    <name>hbase.cluster.distributed</name>
    <value>true</value>
    <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 hbase-env.sh)
    </description>
  </property>
</configuration>

    

    example1
    example3
    example4
    example5
    example6
    example7
    example8
    example9
    

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

 # 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.

    

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

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.

In 0.92.0, the hbase.hregion.memstore.mslab.enabled flag is set to true (See Section 11.3.1.1, “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.

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.

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.

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).

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).

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).

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.

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

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 hbase-env.sh 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: http://www.cloudera.com/blog/2012/01/caching-in-hbase-slabcache/

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.

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’)

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”).

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.

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> SimpleDateFormat.new("yy/MM/dd HH:mm:ss").parse("08/08/16 20:56:29", ParsePosition.new(0)).getTime() => 1218920189000

To go the other direction:

                    hbase(main):021:0> import java.util.Date
                    hbase(main):022:0> Date.new(1218920189000).toString() => "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.

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

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).

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".

HTable htable = ...      // instantiate HTable

Scan scan = new Scan();
scan.addColumn(Bytes.toBytes("cf"),Bytes.toBytes("attr"));
scan.setStartRow( Bytes.toBytes("row"));                   // start key is inclusive
scan.setStopRow( Bytes.toBytes("row" +  (char)0));  // stop key is exclusive
ResultScanner rs = htable.getScanner(scan);
try {
  for (Result r = rs.next(); r != null; r = rs.next()) {
  // process result...
} finally {
  rs.close();  // always close the ResultScanner!
}

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 5.8.1.5, “Delete” for more information on deleting versions of columns, and see Section 9.7.5.5, “Compaction” for more information on compactions.

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

Get get = new Get(Bytes.toBytes("row1"));
Result r = htable.get(get);
byte[] b = r.getValue(Bytes.toBytes("cf"), Bytes.toBytes("attr"));  // returns current version of value

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(Bytes.toBytes("cf"), Bytes.toBytes("attr"));  // returns current version of value
List<KeyValue> kv = r.getColumn(Bytes.toBytes("cf"), Bytes.toBytes("attr"));  // returns all versions of this column

Put put = new Put(Bytes.toBytes(row));
put.add(Bytes.toBytes("cf"), Bytes.toBytes("attr1"), Bytes.toBytes( data));
htable.put(put);

Put put = new Put( Bytes.toBytes(row));
long explicitTimeInMs = 555;  // just an example
put.add(Bytes.toBytes("cf"), Bytes.toBytes("attr1"), explicitTimeInMs, Bytes.toBytes(data));
htable.put(put);

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 9.7.5, “Store” for more information on StoreFiles.

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.

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.

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 9.7.5.2, “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 9.7.5.4, “KeyValue” for more information on HBase stores data internally to see why this is important.

// 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

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.

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

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).

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) {
    logger.info("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;
}

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.

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.

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.

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).

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.

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 ”).

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.

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

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).

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."

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 Lessons Learned from OpenTSDB from HBaseCon2012.

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.

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

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
...

TableMapReduceUtil.initTableMapperJob(
  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);
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.
   }
}
    

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

TableMapReduceUtil.initTableMapperJob(
	sourceTable,      // input table
	scan,	          // Scan instance to control CF and attribute selection
	MyMapper.class,   // mapper class
	null,	          // mapper output key
	null,	          // mapper output value
	job);
TableMapReduceUtil.initTableReducerJob(
	targetTable,      // output table
	null,             // reducer class
	job);
job.setNumReduceTasks(0);

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()) {
			put.add(kv);
		}
		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.

TODO: example for MultiTableOutputFormat.

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

TableMapReduceUtil.initTableMapperJob(
	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);
TableMapReduceUtil.initTableReducerJob(
	targetTable,        // output table
	MyTableReducer.class,    // reducer class
	job);
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>  {

	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(Bytes.toBytes("cf"), Bytes.toBytes("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 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(Bytes.toBytes("cf"), Bytes.toBytes("count"), Bytes.toBytes(i));

    		context.write(null, put);
   	}
}
    

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

TableMapReduceUtil.initTableMapperJob(
	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);
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!");
}
    

 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));
	}
}
    

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.

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.

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/fully.qualified.domain.name@YOUR-REALM.COM. 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.

      <property>
        <name>hbase.regionserver.kerberos.principal</name>
        <value>hbase/_HOST@YOUR-REALM.COM</value>
      </property>
      <property>
        <name>hbase.regionserver.keytab.file</name>
        <value>/etc/hbase/conf/keytab.krb5</value>
      </property>
      <property>
        <name>hbase.master.kerberos.principal</name>
        <value>hbase/_HOST@YOUR-REALM.COM</value>
      </property>
      <property>
        <name>hbase.master.keytab.file</name>
        <value>/etc/hbase/conf/keytab.krb5</value>
      </property>
    

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.

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

      <property>
        <name>hbase.security.authentication</name>
        <value>kerberos</value> 
      </property> 
      <property>
        <name>hbase.security.authorization</name>
        <value>true</value>
      </property>
      <property>
      <name>hbase.coprocessor.region.classes</name>
        <value>org.apache.hadoop.hbase.security.token.TokenProvider</value>
      </property>
    

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

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

      <property>
        <name>hbase.security.authentication</name>
        <value>kerberos</value>
      </property> 
    

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 hbase.security.authentication 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:

      <property>
        <name>hbase.rpc.protection</name>
        <value>privacy</value>
      </property>
    

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("hbase.rpc.protection", "privacy");
      HTable table = new HTable(conf, tablename);
    

Expect a ~10% performance penalty for encrypted communication.

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

    <property>
      <name>hbase.thrift.keytab.file</name>
      <value>/etc/hbase/conf/hbase.keytab</value>
    </property>
    <property>
      <name>hbase.thrift.kerberos.principal</name>
      <value>$USER/_HOST@HADOOP.LOCALDOMAIN</value>
    </property>
    

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

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.

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

    <property>
      <name>hbase.rest.keytab.file</name>
      <value>$KEYTAB</value>
    </property>
    <property>
      <name>hbase.rest.kerberos.principal</name>
      <value>$USER/_HOST@HADOOP.LOCALDOMAIN</value>
    </property>
    

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.

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.

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.

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:

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.

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

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.

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:

      <property>
        <name>hbase.coprocessor.master.classes</name>
        <value>org.apache.hadoop.hbase.security.access.AccessController</value>
      </property>
      <property>
      <name>hbase.coprocessor.region.classes</name>
        <value>org.apache.hadoop.hbase.security.token.TokenProvider,
        org.apache.hadoop.hbase.security.access.AccessController</value>
      </property>
    

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

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

Alter

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>
    

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:

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.

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.

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

Key:

Values:

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

Key:

Values:

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.

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”.

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.

// 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
table.close();
// use the connection for other access to the cluster
connection.close();

If Section 11.7.4, “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.

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

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.

Structural Filters contain other Filters.

FilterList list = new FilterList(FilterList.Operator.MUST_PASS_ONE);
SingleColumnValueFilter filter1 = new SingleColumnValueFilter(
	cf,
	column,
	CompareOp.EQUAL,
	Bytes.toBytes("my value")
	);
list.add(filter1);
SingleColumnValueFilter filter2 = new SingleColumnValueFilter(
	cf,
	column,
	CompareOp.EQUAL,
	Bytes.toBytes("my other value")
	);
list.add(filter2);
scan.setFilter(list);

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 9.4.2.1, “SingleColumnValueFilter”.

RegexStringComparator comp = new RegexStringComparator("my.");   // any value that starts with 'my'
SingleColumnValueFilter filter = new SingleColumnValueFilter(
	cf,
	column,
	CompareOp.EQUAL,
	comp
	);
scan.setFilter(filter);

SubstringComparator comp = new SubstringComparator("y val");   // looking for 'my value'
SingleColumnValueFilter filter = new SingleColumnValueFilter(
	cf,
	column,
	CompareOp.EQUAL,
	comp
	);
scan.setFilter(filter);

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.

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.setFilter(f);
scan.setBatch(10); // set this if there could be many columns returned
ResultScanner rs = t.getScanner(scan);
for (Result r = rs.next(); r != null; r = rs.next()) {
  for (KeyValue kv : r.raw()) {
    // each kv represents a column
  }
}
rs.close();

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.setFilter(f);
scan.setBatch(10); // set this if there could be many columns returned
ResultScanner rs = t.getScanner(scan);
for (Result r = rs.next(); r != null; r = rs.next()) {
  for (KeyValue kv : r.raw()) {
    // each kv represents a column
  }
}
rs.close();

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.setFilter(f);
scan.setBatch(10); // set this if there could be many columns returned
ResultScanner rs = t.getScanner(scan);
for (Result r = rs.next(); r != null; r = rs.next()) {
  for (KeyValue kv : r.raw()) {
    // each kv represents a column
  }
}
rs.close();

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.

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.

The methods exposed by HMasterInterface are primarily metadata-oriented methods:

For example, when the HBaseAdmin method disableTable is invoked, it is serviced by the Master server.

The Master runs several background threads:

The methods exposed by HRegionRegionInterface contain both data-oriented and region-maintenance methods:

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.

The RegionServer runs a variety of background threads:

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.

Determining the "right" region size can be tricky, and there are a few factors to consider:

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).

See Section 2.5.2.6, “Bigger Regions” for more information on configuration.

This section describes how Regions are assigned to RegionServers.

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:

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.

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 2.5.2.7, “Managed Splitting” for how to manually manage splits (and for why you might do this)

HTableDescriptor myHtd = ...;
myHtd.setValue(HTableDescriptor.SPLIT_POLICY, MyCustomSplitPolicy.class.getName());

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.

$ ${HBASE_HOME}/bin/hbase org.apache.hadoop.hbase.io.hfile.HFile  

 $ ${HBASE_HOME}/bin/hbase org.apache.hadoop.hbase.io.hfile.HFile -v -f hdfs://10.81.47.41:8020/hbase/TEST/1418428042/DSMP/4759508618286845475  

/* normal skew:
 *
 *         older ----> newer
 *     _
 *    | |   _
 *    | |  | |   _
 *  --|-|- |-|- |-|---_-------_-------  minCompactSize
 *    | |  | |  | |  | |  _  | |
 *    | |  | |  | |  | | | | | |
 *    | |  | |  | |  | | | | | |
 */

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.

The HBase bulk load process consists of two main steps.

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:

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.

For more information about the referenced utilities, see Section 14.1.9, “ImportTsv” and Section 14.1.10, “CompleteBulkLoad”.

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 ImportTsv.java 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.

The NameNode is responsible for maintaining the filesystem metadata. See the above HDFS Architecture link for more information.

The DataNodes are responsible for storing HDFS blocks. See the above HDFS Architecture link for more information.

RAM, RAM, RAM. Don't starve HBase.

Use a 64-bit platform (and 64-bit JVM).

Watch out for swapping. Set swappiness to 0.

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.

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:

Multiple rack configurations carry the same potential issues as multiple switches, and can suffer performance degradation from two main areas:

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.

Are all the network interfaces functioning correctly? Are you sure? See the Troubleshooting Case Study in Section 13.3.1, “Case Study #1 (Performance Issue On A Single Node)”.

The number of regions for an HBase table is driven by the Section 2.5.2.6, “Bigger Regions”. Also, see the architecture section on Section 9.7.1, “Region Size”

For larger systems, managing compactions and splits may be something you want to consider.

See Section 2.5.2.3, “hbase.regionserver.handler.count”.

See ???. 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.

See ???. If there is blocking in the RegionServer logs, increasing this can help.

See ???. If there is enough RAM, increasing this can help.

Have HBase write the checksum into the datablock and save having to do the checksum seek whenever you read. See the release note on HBASE-5074 support checksums in HBase block cache.

See Section 6.2, “ On the number of column families ”.

See Section 6.3.2, “Try to minimize row and column sizes”. See also Section 11.6.7.1, “However...” for compression caveats.

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 11.4.1, “Number of Regions” for more information.

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 11.8.9, “Bloom Filters” for more information or this answer up in quora, How are bloom filters used in HBase?.

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 9.7.5, “Store”for more information.

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.

Production systems should use compression with their ColumnFamily definitions. See Appendix C, Compression In HBase for more information.

Use the bulk load tool if you can. See Section 9.8, “Bulk Loading”. Otherwise, pay attention to the below.

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.

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.

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.

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.

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.

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.

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 11.7.2, “ Table Creation: Pre-Creating Regions ”, as well as Section 11.4, “HBase Configurations”

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.

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.

When columns are selected explicitly with scan.addColumn, HBase will schedule seek operations to seek between the selected columns. When rows have few columns and each column has only a few versions this can be inefficient. A seek operation is generally slower if does not seek at least past 5-10 columns/versions or 512-1024 bytes.

In order to opportunistically look ahead a few columns/versions to see if the next column/version can be found that way before a seek operation is scheduled, a new attribute Scan.HINT_LOOKAHEAD can be set the on Scan object. The following code instructs the RegionServer to attempt two iterations of next before a seek is scheduled:

Scan scan = new Scan();
scan.addColumn(...);
scan.setAttribute(Scan.HINT_LOOKAHEAD, Bytes.toBytes(2));
table.getScanner(scan);

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 13.3.1, “Case Study #1 (Performance Issue On A Single Node)”.

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 = rs.next(); r != null; r = rs.next()) {
  // process result...
} finally {
  rs.close();  // always close the ResultScanner!
}
htable.close();

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.

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.

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 11.7.2, “ Table Creation: Pre-Creating Regions ”, as well as Section 11.4, “HBase Configurations”

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.^[25][26]

See also Section 11.6.4, “Bloom Filters”.

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 9.7.5.5, “Compaction” and HBaseAdmin.majorCompact.

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).

See http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/client/HTable.html#delete%28org.apache.hadoop.hbase.client.Delete%29

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.

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^[27]. 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 dfs.client.read.shortcircuit 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 Section 11.4.9, “hbase.regionserver.checksum.verify”.

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.

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. Not that there isn't room for improvement (and this gap will, over time, be reduced), but HDFS will always be faster in this use-case.

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.

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 hbase-env.sh add:

export HBASE_OPTS="-XX:+UseConcMarkSweepGC -verbose:gc -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -Xloggc:/home/hadoop/hbase/logs/gc-hbase.log"
          

Adjust the log directory to wherever you log. Note: The GC log does NOT roll automatically, so you'll have to keep an eye on it so it doesn't fill up the disk.

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 this to HBASE_OPTS:

export HBASE_OPTS="-XX:NewSize=64m -XX:MaxNewSize=64m <cms options from above> <gc logging options from above>"
            

For more information on GC pauses, see the 3 part blog post by Todd Lipcon and Section 11.3.1.1, “Long GC pauses” above.

search-hadoop.com 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.

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 12.3.1, “search-hadoop.com”. Take some time crafting your question^[28]; 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.

#hbase on irc.freenode.net

JIRA is also really helpful when looking for Hadoop/HBase-specific issues.

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 11.8.1, “Scan Caching”.

In some situations clients that fetch data from a RegionServer get a LeaseException instead of the usual Section 12.5.1, “ScannerTimeoutException or UnknownScannerException”. Usually the source of the exception is org.apache.hadoop.hbase.regionserver.Leases.removeLease(Leases.java:230) (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 > hbase.regionserver.lease.period. Harsh J investigated the issue as part of the mailing list thread HBase, mail # user - Lease does not exist exceptions

Since 0.20.0 the default log level for org.apache.hadoop.hbase.*is DEBUG.

On your clients, edit $HBASE_HOME/conf/log4j.properties and change this: log4j.logger.org.apache.hadoop.hbase=DEBUG to this: log4j.logger.org.apache.hadoop.hbase=INFO, or even log4j.logger.org.apache.hadoop.hbase=WARN.

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 11.7.2, “ 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 11.4, “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.

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
 java.net.ConnectException: Connection refused: no further information
        at sun.nio.ch.SocketChannelImpl.checkConnect(Native Method)
        at sun.nio.ch.SocketChannelImpl.finishConnect(Unknown Source)
        at org.apache.zookeeper.ClientCnxn$SendThread.run(ClientCnxn.java:1078)
 11/07/05 11:26:43 INFO zookeeper.ClientCnxn: Opening socket connection to
 server localhost/127.0.0.1:2181
 11/07/05 11:26:44 WARN zookeeper.ClientCnxn: Session 0x0 for server null,
 unexpected error, closing socket connection and attempting reconnect
 java.net.ConnectException: Connection refused: no further information
        at sun.nio.ch.SocketChannelImpl.checkConnect(Native Method)
        at sun.nio.ch.SocketChannelImpl.finishConnect(Unknown Source)
        at org.apache.zookeeper.ClientCnxn$SendThread.run(ClientCnxn.java:1078)
 11/07/05 11:26:45 INFO zookeeper.ClientCnxn: Opening socket connection to
 server localhost/127.0.0.1:2181

... are either due to ZooKeeper being down, or unreachable due to network issues.

The utility Section 12.4.1.3, “zkcli” may help investigate ZooKeeper issues.

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).

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.

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 javax.security.auth.useSubjectCredsOnly 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.

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(TotalOrderPartitioner.java:111)
       at org.apache.hadoop.util.ReflectionUtils.setConf(ReflectionUtils.java:62)
       at org.apache.hadoop.util.ReflectionUtils.newInstance(ReflectionUtils.java:117)
       at org.apache.hadoop.mapred.MapTask$NewOutputCollector.<init>(MapTask.java:560)
       at org.apache.hadoop.mapred.MapTask.runNewMapper(MapTask.java:639)
       at org.apache.hadoop.mapred.MapTask.run(MapTask.java:323)
       at org.apache.hadoop.mapred.LocalJobRunner$Job.run(LocalJobRunner.java:210)
Caused by: java.io.FileNotFoundException: File _partition.lst does not exist.
       at org.apache.hadoop.fs.RawLocalFileSystem.getFileStatus(RawLocalFileSystem.java:383)
       at org.apache.hadoop.fs.FilterFileSystem.getFileStatus(FilterFileSystem.java:251)
       at org.apache.hadoop.fs.FileSystem.getLength(FileSystem.java:776)
       at org.apache.hadoop.io.SequenceFile$Reader.<init>(SequenceFile.java:1424)
       at org.apache.hadoop.io.SequenceFile$Reader.<init>(SequenceFile.java:1419)
       at org.apache.hadoop.hbase.mapreduce.hadoopbackport.TotalOrderPartitioner.readPartitions(TotalOrderPartitioner.java:296)

.. see the critical portion of the stack? It's...

       at org.apache.hadoop.mapred.LocalJobRunner$Job.run(LocalJobRunner.java:210)

LocalJobRunner means the job is running locally, not on the cluster.

See http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/mapreduce/package-summary.html#classpath for more information on HBase MapReduce jobs and classpaths.

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.

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...

/hbase
     /<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..

/hbase
     /.logs
          /<RegionServer>    (RegionServers)
               /<HLog>           (WAL HLog files for the RegionServer)
            

See the HDFS User Guide for other non-shell diagnostic utilities like fsck.

If you are seeing periodic network spikes you might want to check the compactionQueues to see if major compactions are happening.

See Section 2.5.2.8, “Managed Compactions” for more information on managing compactions.

HBase expects the loopback IP Address to be 127.0.0.1. See the Getting Started section on Section 2.1.2.3, “Loopback IP”.

Are all the network interfaces functioning correctly? Are you sure? See the Troubleshooting Case Study in Section 12.14, “Case Studies”.

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 12.4.1.3, “zkcli” may help investigate ZooKeeper issues.

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 ec2-174-129-15-236.compute-1.amazonaws.com/10.244.9.171:2181
  2009-10-19 11:52:27,032 WARN org.apache.zookeeper.ClientCnxn: Exception
  closing session 0x0 to sun.nio.ch.SelectionKeyImpl@656dc861
  java.net.ConnectException: 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.

Questions on HBase and Amazon EC2 come up frequently on the HBase dist-list. Search for old threads using Search Hadoop

See Andrew's answer here, up on the user list: Remote Java client connection into EC2 instance.

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>(DefaultMetricsSystem.java:37)
sv4r6s38:       at org.apache.hadoop.metrics2.lib.DefaultMetricsSystem.<clinit>(DefaultMetricsSystem.java:34)
sv4r6s38:       at org.apache.hadoop.security.UgiInstrumentation.create(UgiInstrumentation.java:51)
sv4r6s38:       at org.apache.hadoop.security.UserGroupInformation.initialize(UserGroupInformation.java:209)
sv4r6s38:       at org.apache.hadoop.security.UserGroupInformation.ensureInitialized(UserGroupInformation.java:177)
sv4r6s38:       at org.apache.hadoop.security.UserGroupInformation.isSecurityEnabled(UserGroupInformation.java:229)
sv4r6s38:       at org.apache.hadoop.security.KerberosName.<clinit>(KerberosName.java:83)
sv4r6s38:       at org.apache.hadoop.security.UserGroupInformation.initialize(UserGroupInformation.java:202)
sv4r6s38:       at org.apache.hadoop.security.UserGroupInformation.ensureInitialized(UserGroupInformation.java:177)

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.

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 15.8.3, “Building against various hadoop versions.” for more).

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)
			

<FixedWidthUserName><FixedWidthPageNum0>:<FixedWidthLength><FixedIdNextPageNum><ValueId1><ValueId2><ValueId3>...
<FixedWidthUserName><FixedWidthPageNum1>:<FixedWidthLength><FixedIdNextPageNum><ValueId1><ValueId2><ValueId3>...
    		

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

scan { STARTROW => 'FixedWidthUsername' LIMIT => 30}
    		

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 http://hbase.apache.org/book.html#schema.smackdown).

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: http://www.youtube.com/watch?v=_HLoH_PgrLk).

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! :)