Google Git
Sign in
apache / hbase / 9d71453b085696e8745fec6fbcf9f9c778c77bb3 / . / src / main / docbkx / performance.xml
blob: 58656a71644116b177300e8eba83488d39ff5e71 [file] [log] [blame]
<?xml version="1.0" encoding="UTF-8"?>
<chapter version="5.0" xml:id="performance"
xmlns="http://docbook.org/ns/docbook"
xmlns:xlink="http://www.w3.org/1999/xlink"
xmlns:xi="http://www.w3.org/2001/XInclude"
xmlns:svg="http://www.w3.org/2000/svg"
xmlns:m="http://www.w3.org/1998/Math/MathML"
xmlns:html="http://www.w3.org/1999/xhtml"
xmlns:db="http://docbook.org/ns/docbook">
<!--
/**
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
-->
<title>Apache HBase Performance Tuning</title>
<section xml:id="perf.os">
<title>Operating System</title>
<section xml:id="perf.os.ram">
<title>Memory</title>
<para>RAM, RAM, RAM. Don't starve HBase.</para>
</section>
<section xml:id="perf.os.64">
<title>64-bit</title>
<para>Use a 64-bit platform (and 64-bit JVM).</para>
</section>
<section xml:id="perf.os.swap">
<title>Swapping</title>
<para>Watch out for swapping. Set swappiness to 0.</para>
</section>
</section>
<section xml:id="perf.network">
<title>Network</title>
<para>
Perhaps the most important factor in avoiding network issues degrading Hadoop and HBbase performance is the switching hardware
that is used, decisions made early in the scope of the project can cause major problems when you double or triple the size of your cluster (or more).
</para>
<para>
Important items to consider:
<itemizedlist>
<listitem>Switching capacity of the device</listitem>
<listitem>Number of systems connected</listitem>
<listitem>Uplink capacity</listitem>
</itemizedlist>
</para>
<section xml:id="perf.network.1switch">
<title>Single Switch</title>
<para>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.
</para>
</section>
<section xml:id="perf.network.2switch">
<title>Multiple Switches</title>
<para>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.
</para>
<para>Mitigation of this issue is fairly simple and can be accomplished in multiple ways:
<itemizedlist>
<listitem>Use appropriate hardware for the scale of the cluster which you're attempting to build.</listitem>
<listitem>Use larger single switch configurations i.e. single 48 port as opposed to 2x 24 port</listitem>
<listitem>Configure port trunking for uplinks to utilize multiple interfaces to increase cross switch bandwidth.</listitem>
</itemizedlist>
</para>
</section>
<section xml:id="perf.network.multirack">
<title>Multiple Racks</title>
<para>Multiple rack configurations carry the same potential issues as multiple switches, and can suffer performance degradation from two main areas:
<itemizedlist>
<listitem>Poor switch capacity performance</listitem>
<listitem>Insufficient uplink to another rack</listitem>
</itemizedlist>
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.
</para>
<para>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.
</para>
</section>
<section xml:id="perf.network.ints">
<title>Network Interfaces</title>
<para>Are all the network interfaces functioning correctly? Are you sure? See the Troubleshooting Case Study in <xref linkend="casestudies.slownode"/>.
</para>
</section>
</section> <!-- network -->
<section xml:id="jvm">
<title>Java</title>
<section xml:id="gc">
<title>The Garbage Collector and Apache HBase</title>
<section xml:id="gcpause">
<title>Long GC pauses</title>
<para xml:id="mslab">In his presentation, <link
xlink:href="http://www.slideshare.net/cloudera/hbase-hug-presentation">Avoiding
Full GCs with MemStore-Local Allocation Buffers</link>, Todd Lipcon
describes two cases of stop-the-world garbage collections common in
HBase, especially during loading; CMS failure modes and old generation
heap fragmentation brought. To address the first, start the CMS
earlier than default by adding
<code>-XX:CMSInitiatingOccupancyFraction</code> and setting it down
from defaults. Start at 60 or 70 percent (The lower you bring down the
threshold, the more GCing is done, the more CPU used). To address the
second fragmentation issue, Todd added an experimental facility,
<indexterm><primary>MSLAB</primary></indexterm>, that
must be explicitly enabled in Apache HBase 0.90.x (Its defaulted to be on in
Apache 0.92.x HBase). See <code>hbase.hregion.memstore.mslab.enabled</code>
to true in your <classname>Configuration</classname>. See the cited
slides for background and detail<footnote><para>The latest jvms do better
regards fragmentation so make sure you are running a recent release.
Read down in the message,
<link xlink:href="http://osdir.com/ml/hotspot-gc-use/2011-11/msg00002.html">Identifying concurrent mode failures caused by fragmentation</link>.</para></footnote>.
Be aware that when enabled, each MemStore instance will occupy at least
an MSLAB instance of memory. If you have thousands of regions or lots
of regions each with many column families, this allocation of MSLAB
may be responsible for a good portion of your heap allocation and in
an extreme case cause you to OOME. Disable MSLAB in this case, or
lower the amount of memory it uses or float less regions per server.
</para>
<para>If you have a write-heavy workload, check out
<link xlink:href="https://issues.apache.org/jira/browse/HBASE-8163">HBASE-8163 MemStoreChunkPool: An improvement for JAVA GC when using MSLAB</link>.
It describes configurations to lower the amount of young GC during write-heavy loadings. If you do not have HBASE-8163 installed, and you are
trying to improve your young GC times, one trick to consider -- courtesy of our Liang Xie -- is to set the GC config <varname>-XX:PretenureSizeThreshold</varname> in <filename>hbase-env.sh</filename>
to be just smaller than the size of <varname>hbase.hregion.memstore.mslab.chunksize</varname> so MSLAB allocations happen in the
tenured space directly rather than first in the young gen. You'd do this because these MSLAB allocations are going to likely make it
to the old gen anyways and rather than pay the price of a copies between s0 and s1 in eden space followed by the copy up from
young to old gen after the MSLABs have achieved sufficient tenure, save a bit of YGC churn and allocate in the old gen directly.
</para>
<para>For more information about GC logs, see <xref linkend="trouble.log.gc" />.
</para>
</section>
</section>
</section>
<section xml:id="perf.configurations">
<title>HBase Configurations</title>
<para>See <xref linkend="recommended_configurations" />.</para>
<section xml:id="perf.compactions.and.splits">
<title>Managing Compactions</title>
<para>For larger systems, managing <link
linkend="disable.splitting">compactions and splits</link> may be
something you want to consider.</para>
</section>
<section xml:id="perf.handlers">
<title><varname>hbase.regionserver.handler.count</varname></title>
<para>See <xref linkend="hbase.regionserver.handler.count"/>.
</para>
</section>
<section xml:id="perf.hfile.block.cache.size">
<title><varname>hfile.block.cache.size</varname></title>
<para>See <xref linkend="hfile.block.cache.size"/>.
A memory setting for the RegionServer process.
</para>
</section>
<section xml:id="perf.rs.memstore.upperlimit">
<title><varname>hbase.regionserver.global.memstore.upperLimit</varname></title>
<para>See <xref linkend="hbase.regionserver.global.memstore.upperLimit"/>.
This memory setting is often adjusted for the RegionServer process depending on needs.
</para>
</section>
<section xml:id="perf.rs.memstore.lowerlimit">
<title><varname>hbase.regionserver.global.memstore.lowerLimit</varname></title>
<para>See <xref linkend="hbase.regionserver.global.memstore.lowerLimit"/>.
This memory setting is often adjusted for the RegionServer process depending on needs.
</para>
</section>
<section xml:id="perf.hstore.blockingstorefiles">
<title><varname>hbase.hstore.blockingStoreFiles</varname></title>
<para>See <xref linkend="hbase.hstore.blockingStoreFiles"/>.
If there is blocking in the RegionServer logs, increasing this can help.
</para>
</section>
<section xml:id="perf.hregion.memstore.block.multiplier">
<title><varname>hbase.hregion.memstore.block.multiplier</varname></title>
<para>See <xref linkend="hbase.hregion.memstore.block.multiplier"/>.
If there is enough RAM, increasing this can help.
</para>
</section>
<section xml:id="hbase.regionserver.checksum.verify">
<title><varname>hbase.regionserver.checksum.verify</varname></title>
<para>Have HBase write the checksum into the datablock and save
having to do the checksum seek whenever you read.</para>
<para>See <xref linkend="hbase.regionserver.checksum.verify"/>,
<xref linkend="hbase.hstore.bytes.per.checksum"/> and <xref linkend="hbase.hstore.checksum.algorithm"/>
For more information see the
release note on <link xlink:href="https://issues.apache.org/jira/browse/HBASE-5074">HBASE-5074 support checksums in HBase block cache</link>.
</para>
</section>
</section>
<section xml:id="perf.zookeeper">
<title>ZooKeeper</title>
<para>See <xref linkend="zookeeper"/> for information on configuring ZooKeeper, and see the part
about having a dedicated disk.
</para>
</section>
<section xml:id="perf.schema">
<title>Schema Design</title>
<section xml:id="perf.number.of.cfs">
<title>Number of Column Families</title>
<para>See <xref linkend="number.of.cfs" />.</para>
</section>
<section xml:id="perf.schema.keys">
<title>Key and Attribute Lengths</title>
<para>See <xref linkend="keysize" />. See also <xref linkend="perf.compression.however" /> for
compression caveats.</para>
</section>
<section xml:id="schema.regionsize"><title>Table RegionSize</title>
<para>The regionsize can be set on a per-table basis via <code>setFileSize</code> on
<link xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/HTableDescriptor.html">HTableDescriptor</link> in the
event where certain tables require different regionsizes than the configured default regionsize.
</para>
<para>See <xref linkend="ops.capacity.regions"/> for more information.
</para>
</section>
<section xml:id="schema.bloom">
<title>Bloom Filters</title>
<para>Bloom Filters can be enabled per-ColumnFamily.
Use <code>HColumnDescriptor.setBloomFilterType(NONE | ROW |
ROWCOL)</code> to enable blooms per Column Family. Default =
<varname>NONE</varname> for no bloom filters. If
<varname>ROW</varname>, the hash of the row will be added to the bloom
on each insert. If <varname>ROWCOL</varname>, the hash of the row +
column family + column family qualifier will be added to the bloom on
each key insert.</para>
<para>See <link xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/HColumnDescriptor.html">HColumnDescriptor</link> and
<xref linkend="blooms"/> for more information or this answer up in quora,
<link xlink:href="http://www.quora.com/How-are-bloom-filters-used-in-HBase">How are bloom filters used in HBase?</link>.
</para>
</section>
<section xml:id="schema.cf.blocksize"><title>ColumnFamily BlockSize</title>
<para>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).
</para>
<para>See <link xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/HColumnDescriptor.html">HColumnDescriptor</link>
and <xref linkend="store"/>for more information.
</para>
</section>
<section xml:id="cf.in.memory">
<title>In-Memory ColumnFamilies</title>
<para>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 <xref linkend="block.cache" />, but it is not a guarantee that the entire table
will be in memory.
</para>
<para>See <link xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/HColumnDescriptor.html">HColumnDescriptor</link> for more information.
</para>
</section>
<section xml:id="perf.compression">
<title>Compression</title>
<para>Production systems should use compression with their ColumnFamily definitions. See <xref linkend="compression" /> for more information.
</para>
<section xml:id="perf.compression.however"><title>However...</title>
<para>Compression deflates data <emphasis>on disk</emphasis>. When it's in-memory (e.g., in the
MemStore) or on the wire (e.g., transferring between RegionServer and Client) it's inflated.
So while using ColumnFamily compression is a best practice, but it's not going to completely eliminate
the impact of over-sized Keys, over-sized ColumnFamily names, or over-sized Column names.
</para>
<para>See <xref linkend="keysize" /> on for schema design tips, and <xref linkend="keyvalue"/> for more information on HBase stores data internally.
</para>
</section>
</section>
</section> <!-- perf schema -->
<section xml:id="perf.general">
<title>HBase General Patterns</title>
<section xml:id="perf.general.constants">
<title>Constants</title>
<para>When people get started with HBase they have a tendency to write code that looks like this:
<programlisting>
Get get = new Get(rowkey);
Result r = htable.get(get);
byte[] b = r.getValue(Bytes.toBytes("cf"), Bytes.toBytes("attr")); // returns current version of value
</programlisting>
But especially when inside loops (and MapReduce jobs), converting the columnFamily and column-names
to byte-arrays repeatedly is surprisingly expensive.
It's better to use constants for the byte-arrays, like this:
<programlisting>
public static final byte[] CF = "cf".getBytes();
public static final byte[] ATTR = "attr".getBytes();
...
Get get = new Get(rowkey);
Result r = htable.get(get);
byte[] b = r.getValue(CF, ATTR); // returns current version of value
</programlisting>
</para>
</section>
</section>
<section xml:id="perf.writing">
<title>Writing to HBase</title>
<section xml:id="perf.batch.loading">
<title>Batch Loading</title>
<para>Use the bulk load tool if you can. See
<xref linkend="arch.bulk.load"/>.
Otherwise, pay attention to the below.
</para>
</section> <!-- batch loading -->
<section xml:id="precreate.regions">
<title>
Table Creation: Pre-Creating Regions
</title>
<para>
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.
</para>
<para>There are two different approaches to pre-creating splits. The first approach is to rely on the default <code>HBaseAdmin</code> strategy
(which is implemented in <code>Bytes.split</code>)...
</para>
<programlisting>
byte[] startKey = ...; // your lowest keuy
byte[] endKey = ...; // your highest key
int numberOfRegions = ...; // # of regions to create
admin.createTable(table, startKey, endKey, numberOfRegions);
</programlisting>
<para>And the other approach is to define the splits yourself...
</para>
<programlisting>
byte[][] splits = ...; // create your own splits
admin.createTable(table, splits);
</programlisting>
<para>
See <xref linkend="rowkey.regionsplits"/> for issues related to understanding your keyspace and pre-creating regions.
</para>
</section>
<section xml:id="def.log.flush">
<title>
Table Creation: Deferred Log Flush
</title>
<para>
The default behavior for Puts using the Write Ahead Log (WAL) is that <classname>HLog</classname> 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 <classname>HLog</classname>- 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.
</para>
<para>
Deferred log flush can be configured on tables via <link
xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/HTableDescriptor.html">HTableDescriptor</link>. The default value of <varname>hbase.regionserver.optionallogflushinterval</varname> is 1000ms.
</para>
</section>
<section xml:id="perf.hbase.client.autoflush">
<title>HBase Client: AutoFlush</title>
<para>When performing a lot of Puts, make sure that setAutoFlush is set
to false on your <link
xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/client/HTable.html">HTable</link>
instance. Otherwise, the Puts will be sent one at a time to the
RegionServer. Puts added via <code> htable.add(Put)</code> and <code> htable.add( &lt;List&gt; Put)</code>
wind up in the same write buffer. If <code>autoFlush = false</code>,
these messages are not sent until the write-buffer is filled. To
explicitly flush the messages, call <methodname>flushCommits</methodname>.
Calling <methodname>close</methodname> on the <classname>HTable</classname>
instance will invoke <methodname>flushCommits</methodname>.</para>
</section>
<section xml:id="perf.hbase.client.putwal">
<title>HBase Client: Turn off WAL on Puts</title>
<para>A frequently discussed option for increasing throughput on <classname>Put</classname>s is to call <code>writeToWAL(false)</code>. Turning this off means
that the RegionServer will <emphasis>not</emphasis> write the <classname>Put</classname> to the Write Ahead Log,
only into the memstore, HOWEVER the consequence is that if there
is a RegionServer failure <emphasis>there will be data loss</emphasis>.
If <code>writeToWAL(false)</code> 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.
</para>
<para>In general, it is best to use WAL for Puts, and where loading throughput
is a concern to use <link linkend="perf.batch.loading">bulk loading</link> techniques instead.
</para>
</section>
<section xml:id="perf.hbase.client.regiongroup">
<title>HBase Client: Group Puts by RegionServer</title>
<para>In addition to using the writeBuffer, grouping <classname>Put</classname>s by RegionServer can reduce the number of client RPC calls per writeBuffer flush.
There is a utility <classname>HTableUtil</classname> 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.
</para>
</section>
<section xml:id="perf.hbase.write.mr.reducer">
<title>MapReduce: Skip The Reducer</title>
<para>When writing a lot of data to an HBase table from a MR job (e.g., with <link
xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/mapreduce/TableOutputFormat.html">TableOutputFormat</link>), 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.
</para>
<para>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.
</para>
</section>
<section xml:id="perf.one.region">
<title>Anti-Pattern: One Hot Region</title>
<para>If all your data is being written to one region at a time, then re-read the
section on processing <link linkend="timeseries">timeseries</link> data.</para>
<para>Also, if you are pre-splitting regions and all your data is <emphasis>still</emphasis> 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
<link xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/client/HTable.html#getRegionLocation%28byte[]%29">HTable.getRegionLocation</link>.
</para>
<para>See <xref linkend="precreate.regions"/>, as well as <xref linkend="perf.configurations"/> </para>
</section>
</section> <!-- writing -->
<section xml:id="perf.reading">
<title>Reading from HBase</title>
<section xml:id="perf.hbase.client.caching">
<title>Scan Caching</title>
<para>If HBase is used as an input source for a MapReduce job, for
example, make sure that the input <link
xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/client/Scan.html">Scan</link>
instance to the MapReduce job has <methodname>setCaching</methodname> 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.</para>
<section xml:id="perf.hbase.client.caching.mr">
<title>Scan Caching in MapReduce Jobs</title>
<para>Scan settings in MapReduce jobs deserve special attention. Timeouts can result (e.g., UnknownScannerException)
in Map tasks if it takes longer to process a batch of records before the client goes back to the RegionServer for the
next set of data. This problem can occur because there is non-trivial processing occuring per row. If you process
rows quickly, set caching higher. If you process rows more slowly (e.g., lots of transformations per row, writes),
then set caching lower.
</para>
<para>Timeouts can also happen in a non-MapReduce use case (i.e., single threaded HBase client doing a Scan), but the
processing that is often performed in MapReduce jobs tends to exacerbate this issue.
</para>
</section>
</section>
<section xml:id="perf.hbase.client.selection">
<title>Scan Attribute Selection</title>
<para>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 <code>scan.addFamily</code> is called
then <emphasis>all</emphasis> 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.
</para>
</section>
<section xml:id="perf.hbase.mr.input">
<title>MapReduce - Input Splits</title>
<para>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 <xref linkend="casestudies.slownode"/>.
</para>
</section>
<section xml:id="perf.hbase.client.scannerclose">
<title>Close ResultScanners</title>
<para>This isn't so much about improving performance but rather
<emphasis>avoiding</emphasis> performance problems. If you forget to
close <link
xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/client/ResultScanner.html">ResultScanners</link>
you can cause problems on the RegionServers. Always have ResultScanner
processing enclosed in try/catch blocks... <programlisting>
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();</programlisting></para>
</section>
<section xml:id="perf.hbase.client.blockcache">
<title>Block Cache</title>
<para><link
xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/client/Scan.html">Scan</link>
instances can be set to use the block cache in the RegionServer via the
<methodname>setCacheBlocks</methodname> method. For input Scans to MapReduce jobs, this should be
<varname>false</varname>. For frequently accessed rows, it is advisable to use the block
cache.</para>
</section>
<section xml:id="perf.hbase.client.rowkeyonly">
<title>Optimal Loading of Row Keys</title>
<para>When performing a table <link xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/client/Scan.html">scan</link>
where only the row keys are needed (no families, qualifiers, values or timestamps), add a FilterList with a
<varname>MUST_PASS_ALL</varname> operator to the scanner using <methodname>setFilter</methodname>. The filter list
should include both a <link xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/filter/FirstKeyOnlyFilter.html">FirstKeyOnlyFilter</link>
and a <link xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/filter/KeyOnlyFilter.html">KeyOnlyFilter</link>.
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.
</para>
</section>
<section xml:id="perf.hbase.read.dist">
<title>Concurrency: Monitor Data Spread</title>
<para>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. </para>
<para>See <xref linkend="precreate.regions"/>, as well as <xref linkend="perf.configurations"/> </para>
</section>
<section xml:id="blooms">
<title>Bloom Filters</title>
<para>Enabling Bloom Filters can save your having to go to disk and
can help improve read latencys.</para>
<para><link xlink:href="http://en.wikipedia.org/wiki/Bloom_filter">Bloom filters</link> were developed over in <link
xlink:href="https://issues.apache.org/jira/browse/HBASE-1200">HBase-1200
Add bloomfilters</link>.<footnote>
<para>For description of the development process -- why static blooms
rather than dynamic -- and for an overview of the unique properties
that pertain to blooms in HBase, as well as possible future
directions, see the <emphasis>Development Process</emphasis> section
of the document <link
xlink:href="https://issues.apache.org/jira/secure/attachment/12444007/Bloom_Filters_in_HBase.pdf">BloomFilters
in HBase</link> attached to <link
xlink:href="https://issues.apache.org/jira/browse/HBASE-1200">HBase-1200</link>.</para>
</footnote><footnote>
<para>The bloom filters described here are actually version two of
blooms in HBase. In versions up to 0.19.x, HBase had a dynamic bloom
option based on work done by the <link
xlink:href="http://www.one-lab.org">European Commission One-Lab
Project 034819</link>. The core of the HBase bloom work was later
pulled up into Hadoop to implement org.apache.hadoop.io.BloomMapFile.
Version 1 of HBase blooms never worked that well. Version 2 is a
rewrite from scratch though again it starts with the one-lab
work.</para>
</footnote></para>
<para>See also <xref linkend="schema.bloom" />.
</para>
<section xml:id="bloom_footprint">
<title>Bloom StoreFile footprint</title>
<para>Bloom filters add an entry to the <classname>StoreFile</classname>
general <classname>FileInfo</classname> data structure and then two
extra entries to the <classname>StoreFile</classname> metadata
section.</para>
<section>
<title>BloomFilter in the <classname>StoreFile</classname>
<classname>FileInfo</classname> data structure</title>
<para><classname>FileInfo</classname> has a
<varname>BLOOM_FILTER_TYPE</varname> entry which is set to
<varname>NONE</varname>, <varname>ROW</varname> or
<varname>ROWCOL.</varname></para>
</section>
<section>
<title>BloomFilter entries in <classname>StoreFile</classname>
metadata</title>
<para><varname>BLOOM_FILTER_META</varname> holds Bloom Size, Hash
Function used, etc. Its small in size and is cached on
<classname>StoreFile.Reader</classname> load</para>
<para><varname>BLOOM_FILTER_DATA</varname> is the actual bloomfilter
data. Obtained on-demand. Stored in the LRU cache, if it is enabled
(Its enabled by default).</para>
</section>
</section>
<section xml:id="config.bloom">
<title>Bloom Filter Configuration</title>
<section>
<title><varname>io.hfile.bloom.enabled</varname> global kill
switch</title>
<para><code>io.hfile.bloom.enabled</code> in
<classname>Configuration</classname> serves as the kill switch in case
something goes wrong. Default = <varname>true</varname>.</para>
</section>
<section>
<title><varname>io.hfile.bloom.error.rate</varname></title>
<para><varname>io.hfile.bloom.error.rate</varname> = average false
positive rate. Default = 1%. Decrease rate by ½ (e.g. to .5%) == +1
bit per bloom entry.</para>
</section>
<section>
<title><varname>io.hfile.bloom.max.fold</varname></title>
<para><varname>io.hfile.bloom.max.fold</varname> = guaranteed minimum
fold rate. Most people should leave this alone. Default = 7, or can
collapse to at least 1/128th of original size. See the
<emphasis>Development Process</emphasis> section of the document <link
xlink:href="https://issues.apache.org/jira/secure/attachment/12444007/Bloom_Filters_in_HBase.pdf">BloomFilters
in HBase</link> for more on what this option means.</para>
</section>
</section>
</section> <!-- bloom -->
</section> <!-- reading -->
<section xml:id="perf.deleting">
<title>Deleting from HBase</title>
<section xml:id="perf.deleting.queue">
<title>Using HBase Tables as Queues</title>
<para>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 <xref linkend="datamodel" />, marking rows as deleted creates additional StoreFiles which then need to be processed
on reads. Tombstones only get cleaned up with major compactions.
</para>
<para>See also <xref linkend="compaction" /> and <link xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/client/HBaseAdmin.html#majorCompact%28java.lang.String%29">HBaseAdmin.majorCompact</link>.
</para>
</section>
<section xml:id="perf.deleting.rpc">
<title>Delete RPC Behavior</title>
<para>Be aware that <code>htable.delete(Delete)</code> doesn't use the writeBuffer. It will execute an RegionServer RPC with each invocation.
For a large number of deletes, consider <code>htable.delete(List)</code>.
</para>
<para>See <link xlink:href="http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/client/HTable.html#delete%28org.apache.hadoop.hbase.client.Delete%29"></link>
</para>
</section>
</section> <!-- deleting -->
<section xml:id="perf.hdfs"><title>HDFS</title>
<para>Because HBase runs on <xref linkend="arch.hdfs" /> it is important to understand how it works and how it affects
HBase.
</para>
<section xml:id="perf.hdfs.curr"><title>Current Issues With Low-Latency Reads</title>
<para>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
<link xlink:href="https://issues.apache.org/jira/browse/HDFS-1599">Umbrella Jira Ticket for HDFS Improvements for HBase</link>.
</para>
</section>
<section xml:id="perf.hdfs.configs.localread">
<title>Leveraging local data</title>
<para>Since Hadoop 1.0.0 (also 0.22.1, 0.23.1, CDH3u3 and HDP 1.0) via
<link xlink:href="https://issues.apache.org/jira/browse/HDFS-2246">HDFS-2246</link>,
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<footnote><para>See JD's <link xlink:href="http://files.meetup.com/1350427/hug_ebay_jdcryans.pdf">Performance Talk</link></para></footnote>.
Also see <link xlink:href="http://search-hadoop.com/m/zV6dKrLCVh1">HBase, mail # dev - read short circuit</link> thread for
more discussion around short circuit reads.
</para>
<para>To enable "short circuit" reads, it will depend on your version of Hadoop.
The original shortcircuit read patch was much improved upon in Hadoop 2 in
<link xlink:href="https://issues.apache.org/jira/browse/HDFS-347">HDFS-347</link>.
See <link xlink:href="http://blog.cloudera.com/blog/2013/08/how-improved-short-circuit-local-reads-bring-better-performance-and-security-to-hadoop/"></link> for details
on the difference between the old and new implementations. See
<link xlink:href="http://archive.cloudera.com/cdh4/cdh/4/hadoop/hadoop-project-dist/hadoop-hdfs/ShortCircuitLocalReads.html">Hadoop shortcircuit reads configuration page</link>
for how to enable the later version of shortcircuit.
</para>
<para>If you are running on an old Hadoop, one that is without
<link xlink:href="https://issues.apache.org/jira/browse/HDFS-347">HDFS-347</link> but that
has
<link xlink:href="https://issues.apache.org/jira/browse/HDFS-2246">HDFS-2246</link>,
you must set two configurations.
First, the hdfs-site.xml needs to be amended. Set
the property <varname>dfs.block.local-path-access.user</varname>
to be the <emphasis>only</emphasis> user that can use the shortcut.
This has to be the user that started HBase. Then in hbase-site.xml,
set <varname>dfs.client.read.shortcircuit</varname> to be <varname>true</varname>
</para>
<para>
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 <xref linkend="hbase.regionserver.checksum.verify" />. When both
local short-circuit reads and hbase level checksums are enabled, you SHOULD NOT disable configuration parameter
"dfs.client.read.shortcircuit.skip.checksum", which will cause skipping checksum on non-hfile reads. HBase already
manages that setting under the covers.
</para>
<para>
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.
</para>
<note xml:id="dfs.client.read.shortcircuit.buffer.size">
<title>dfs.client.read.shortcircuit.buffer.size</title>
<para>The default for this value is too high when running on a highly trafficed HBase. Set it down from its
1M default down to 128k or so. Put this configuration in the HBase configs (its a HDFS client-side configuration).
The Hadoop DFSClient in HBase will allocate a direct byte buffer of this size for <emphasis>each</emphasis>
block it has open; given HBase keeps its HDFS files open all the time, this can add up quickly.</para>
</note>
</section>
<section xml:id="perf.hdfs.comp"><title>Performance Comparisons of HBase vs. HDFS</title>
<para>A fairly common question on the dist-list is why HBase isn't as performant as HDFS files in a batch context (e.g., as
a MapReduce source or sink). The short answer is that HBase is doing a lot more than HDFS (e.g., reading the KeyValues,
returning the most current row or specified timestamps, etc.), and as such HBase is 4-5 times slower than HDFS in this
processing context. There is room for improvement and this gap will, over time, be reduced, but HDFS
will always be faster in this use-case.
</para>
</section>
</section>
<section xml:id="perf.ec2"><title>Amazon EC2</title>
<para>Performance questions are common on Amazon EC2 environments because it is a shared environment. You will
not see the same throughput as a dedicated server. In terms of running tests on EC2, run them several times for the same
reason (i.e., it's a shared environment and you don't know what else is happening on the server).
</para>
<para>If you are running on EC2 and post performance questions on the dist-list, please state this fact up-front that
because EC2 issues are practically a separate class of performance issues.
</para>
</section>
<section xml:id="perf.hbase.mr.cluster"><title>Collocating HBase and MapReduce</title>
<para>It is often recommended to have different clusters for HBase and MapReduce. A better qualification of this is:
don't collocate a HBase that serves live requests with a heavy MR workload. OLTP and OLAP-optimized systems have
conflicting requirements and one will lose to the other, usually the former. For example, short latency-sensitive
disk reads will have to wait in line behind longer reads that are trying to squeeze out as much throughput as
possible. MR jobs that write to HBase will also generate flushes and compactions, which will in turn invalidate
blocks in the <xref linkend="block.cache"/>.
</para>
<para>If you need to process the data from your live HBase cluster in MR, you can ship the deltas with <xref linkend="copy.table"/>
or use replication to get the new data in real time on the OLAP cluster. In the worst case, if you really need to
collocate both, set MR to use less Map and Reduce slots than you'd normally configure, possibly just one.
</para>
<para>When HBase is used for OLAP operations, it's preferable to set it up in a hardened way like configuring the ZooKeeper session
timeout higher and giving more memory to the MemStores (the argument being that the Block Cache won't be used much
since the workloads are usually long scans).
</para>
</section>
<section xml:id="perf.casestudy"><title>Case Studies</title>
<para>For Performance and Troubleshooting Case Studies, see <xref linkend="casestudies"/>.
</para>
</section>
</chapter>