docs/topics/impala_perf_cookbook.xml - impala - Git at Google

 <?xml version="1.0" encoding="UTF-8"?>
 <!--
 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.
 -->
 <!DOCTYPE concept PUBLIC "-//OASIS//DTD DITA Concept//EN" "concept.dtd">
 <concept id="perf_cookbook">

   <title>Impala Performance Guidelines and Best Practices</title>
   <titlealts audience="PDF"><navtitle>Performance Best Practices</navtitle></titlealts>
   <prolog>
     <metadata>
       <data name="Category" value="Performance"/>
       <data name="Category" value="Impala"/>
       <data name="Category" value="Planning"/>
       <data name="Category" value="Proof of Concept"/>
       <data name="Category" value="Guidelines"/>
       <data name="Category" value="Best Practices"/>
       <data name="Category" value="Proof of Concept"/>
       <data name="Category" value="Developers"/>
       <data name="Category" value="Data Analysts"/>
     </metadata>
   </prolog>

   <conbody>

     <p>
       Here are performance guidelines and best practices that you can use during planning, experimentation, and
       performance tuning for an Impala-enabled <keyword keyref="distro"/> cluster. All of this information is also available in more
       detail elsewhere in the Impala documentation; it is gathered together here to serve as a cookbook and
       emphasize which performance techniques typically provide the highest return on investment
     </p>

     <p outputclass="toc inpage"/>

     <section id="perf_cookbook_file_format">

       <title>Choose the appropriate file format for the data</title>

       <p>
         Typically, for large volumes of data (multiple gigabytes per table or partition), the Parquet file format
         performs best because of its combination of columnar storage layout, large I/O request size, and
         compression and encoding. See <xref href="impala_file_formats.xml#file_formats"/> for comparisons of all
         file formats supported by Impala, and <xref href="impala_parquet.xml#parquet"/> for details about the
         Parquet file format.
       </p>

       <note>
         For smaller volumes of data, a few gigabytes or less for each table or partition, you might not see
         significant performance differences between file formats. At small data volumes, reduced I/O from an
         efficient compressed file format can be counterbalanced by reduced opportunity for parallel execution. When
         planning for a production deployment or conducting benchmarks, always use realistic data volumes to get a
         true picture of performance and scalability.
       </note>
     </section>

     <section id="perf_cookbook_small_files">

       <title>Avoid data ingestion processes that produce many small files</title>

       <p>
         When producing data files outside of Impala, prefer either text format or Avro, where you can build up the
         files row by row. Once the data is in Impala, you can convert it to the more efficient Parquet format and
         split into multiple data files using a single <codeph>INSERT ... SELECT</codeph> statement. Or, if you have
         the infrastructure to produce multi-megabyte Parquet files as part of your data preparation process, do
         that and skip the conversion step inside Impala.
       </p>

       <p>
         Always use <codeph>INSERT ... SELECT</codeph> to copy significant volumes of data from table to table
         within Impala. Avoid <codeph>INSERT ... VALUES</codeph> for any substantial volume of data or
         performance-critical tables, because each such statement produces a separate tiny data file. See
         <xref href="impala_insert.xml#insert"/> for examples of the <codeph>INSERT ... SELECT</codeph> syntax.
       </p>

       <p>
         For example, if you have thousands of partitions in a Parquet table, each with less than
         <ph rev="parquet_block_size">256 MB</ph> of data, consider partitioning in a less granular way, such as by
         year / month rather than year / month / day. If an inefficient data ingestion process produces thousands of
         data files in the same table or partition, consider compacting the data by performing an <codeph>INSERT ...
         SELECT</codeph> to copy all the data to a different table; the data will be reorganized into a smaller
         number of larger files by this process.
       </p>
     </section>

     <section id="perf_cookbook_partitioning">

       <title>Choose partitioning granularity based on actual data volume</title>

       <p>
         Partitioning is a technique that physically divides the data based on values of one or more columns, such
         as by year, month, day, region, city, section of a web site, and so on. When you issue queries that request
         a specific value or range of values for the partition key columns, Impala can avoid reading the irrelevant
         data, potentially yielding a huge savings in disk I/O.
       </p>

       <p>
         When deciding which column(s) to use for partitioning, choose the right level of granularity. For example,
         should you partition by year, month, and day, or only by year and month? Choose a partitioning strategy
         that puts at least <ph rev="parquet_block_size">256 MB</ph> of data in each partition, to take advantage of
         HDFS bulk I/O and Impala distributed queries.
       </p>

       <p>
         Over-partitioning can also cause query planning to take longer than necessary, as Impala prunes the
         unnecessary partitions. Ideally, keep the number of partitions in the table under 30 thousand.
       </p>

       <p>
         When preparing data files to go in a partition directory, create several large files rather than many small
         ones. If you receive data in the form of many small files and have no control over the input format,
         consider using the <codeph>INSERT ... SELECT</codeph> syntax to copy data from one table or partition to
         another, which compacts the files into a relatively small number (based on the number of nodes in the
         cluster).
       </p>

       <p>
         If you need to reduce the overall number of partitions and increase the amount of data in each partition,
         first look for partition key columns that are rarely referenced or are referenced in non-critical queries
         (not subject to an SLA). For example, your web site log data might be partitioned by year, month, day, and
         hour, but if most queries roll up the results by day, perhaps you only need to partition by year, month,
         and day.
       </p>

       <p>
         If you need to reduce the granularity even more, consider creating <q>buckets</q>, computed values
         corresponding to different sets of partition key values. For example, you can use the
         <codeph>TRUNC()</codeph> function with a <codeph>TIMESTAMP</codeph> column to group date and time values
         based on intervals such as week or quarter. See
         <xref href="impala_datetime_functions.xml#datetime_functions"/> for details.
       </p>

       <p>
         See <xref href="impala_partitioning.xml#partitioning"/> for full details and performance considerations for
         partitioning.
       </p>
     </section>

     <section id="perf_cookbook_partition_keys">

       <title>Use smallest appropriate integer types for partition key columns</title>

       <p>
         Although it is tempting to use strings for partition key columns, since those values are turned into HDFS
         directory names anyway, you can minimize memory usage by using numeric values for common partition key
         fields such as <codeph>YEAR</codeph>, <codeph>MONTH</codeph>, and <codeph>DAY</codeph>. Use the smallest
         integer type that holds the appropriate range of values, typically <codeph>TINYINT</codeph> for
         <codeph>MONTH</codeph> and <codeph>DAY</codeph>, and <codeph>SMALLINT</codeph> for <codeph>YEAR</codeph>.
         Use the <codeph>EXTRACT()</codeph> function to pull out individual date and time fields from a
         <codeph>TIMESTAMP</codeph> value, and <codeph>CAST()</codeph> the return value to the appropriate integer
         type.
       </p>
     </section>

     <section id="perf_cookbook_parquet_block_size">

       <title>Choose an appropriate Parquet block size</title>

       <p rev="parquet_block_size">
         By default, the Impala <codeph>INSERT ... SELECT</codeph> statement creates Parquet files with a 256 MB
         block size. (This default was changed in Impala 2.0. Formerly, the limit was 1 GB, but Impala made
         conservative estimates about compression, resulting in files that were smaller than 1 GB.)
       </p>

       <p>
         Each Parquet file written by Impala is a single block, allowing the whole file to be processed as a unit by a single host.
         As you copy Parquet files into HDFS or between HDFS filesystems, use <codeph>hdfs dfs -pb</codeph> to preserve the original
         block size.
       </p>

       <p>
         If there is only one or a few data block in your Parquet table, or in a partition that is the only one
         accessed by a query, then you might experience a slowdown for a different reason: not enough data to take
         advantage of Impala's parallel distributed queries. Each data block is processed by a single core on one of
         the DataNodes. In a 100-node cluster of 16-core machines, you could potentially process thousands of data
         files simultaneously. You want to find a sweet spot between <q>many tiny files</q> and <q>single giant
         file</q> that balances bulk I/O and parallel processing. You can set the <codeph>PARQUET_FILE_SIZE</codeph>
         query option before doing an <codeph>INSERT ... SELECT</codeph> statement to reduce the size of each
         generated Parquet file. <ph rev="2.0.0">(Specify the file size as an absolute number of bytes, or in Impala
         2.0 and later, in units ending with <codeph>m</codeph> for megabytes or <codeph>g</codeph> for
         gigabytes.)</ph> Run benchmarks with different file sizes to find the right balance point for your
         particular data volume.
       </p>
     </section>

     <section id="perf_cookbook_stats">

       <title>Gather statistics for all tables used in performance-critical or high-volume join queries</title>

       <p>
         Gather the statistics with the <codeph>COMPUTE STATS</codeph> statement. See
         <xref href="impala_perf_joins.xml#perf_joins"/> for details.
       </p>
     </section>

     <section id="perf_cookbook_network">

       <title>Minimize the overhead of transmitting results back to the client</title>

       <p>
         Use techniques such as:
       </p>

       <ul>
         <li>
           Aggregation. If you need to know how many rows match a condition, the total values of matching values
           from some column, the lowest or highest matching value, and so on, call aggregate functions such as
           <codeph>COUNT()</codeph>, <codeph>SUM()</codeph>, and <codeph>MAX()</codeph> in the query rather than
           sending the result set to an application and doing those computations there. Remember that the size of an
           unaggregated result set could be huge, requiring substantial time to transmit across the network.
         </li>

         <li>
           Filtering. Use all applicable tests in the <codeph>WHERE</codeph> clause of a query to eliminate rows
           that are not relevant, rather than producing a big result set and filtering it using application logic.
         </li>

         <li>
           <codeph>LIMIT</codeph> clause. If you only need to see a few sample values from a result set, or the top
           or bottom values from a query using <codeph>ORDER BY</codeph>, include the <codeph>LIMIT</codeph> clause
           to reduce the size of the result set rather than asking for the full result set and then throwing most of
           the rows away.
         </li>

         <li>
           Avoid overhead from pretty-printing the result set and displaying it on the screen. When you retrieve the
           results through <cmdname>impala-shell</cmdname>, use <cmdname>impala-shell</cmdname> options such as
           <codeph>-B</codeph> and <codeph>--output_delimiter</codeph> to produce results without special
           formatting, and redirect output to a file rather than printing to the screen. Consider using
           <codeph>INSERT ... SELECT</codeph> to write the results directly to new files in HDFS. See
           <xref href="impala_shell_options.xml#shell_options"/> for details about the
           <cmdname>impala-shell</cmdname> command-line options.
         </li>
       </ul>
     </section>

     <section id="perf_cookbook_explain">

       <title>Verify that your queries are planned in an efficient logical manner</title>

       <p>
         Examine the <codeph>EXPLAIN</codeph> plan for a query before actually running it. See
         <xref href="impala_explain.xml#explain"/> and <xref href="impala_explain_plan.xml#perf_explain"/> for
         details.
       </p>
     </section>

     <section id="perf_cookbook_profile">

       <title>Verify performance characteristics of queries</title>

       <p>
         Verify that the low-level aspects of I/O, memory usage, network bandwidth, CPU utilization, and so on are
         within expected ranges by examining the query profile for a query after running it. See
         <xref href="impala_explain_plan.xml#perf_profile"/> for details.
       </p>
     </section>

     <section id="perf_cookbook_os">

       <title>Use appropriate operating system settings</title>

       <p>
         See <xref keyref="cdh_admin_performance"/> for recommendations about operating system
         settings that you can change to influence Impala performance. In particular, you might find
         that changing the <codeph>vm.swappiness</codeph> Linux kernel setting to a non-zero value improves
         overall performance.
       </p>
     </section>
     <section id="perf_cookbook_hotspot">
       <title>Hotspot analysis</title>

       <p>
         In the context of Impala, a hotspot is defined as “an Impala daemon
         that for a single query or a workload is spending a far greater amount
         of time processing data relative to its neighbours”.
       </p>

       <p>
         Before discussing the options to tackle this issue some background is
         first required to understand how this problem can occur.
       </p>

       <p>
         By default, the scheduling of scan based plan fragments is
         deterministic. This means that for multiple queries needing to read the
         same block of data, the same node will be picked to host the scan. The
         default scheduling logic does not take into account node workload from
         prior queries. The complexity of materializing a tuple depends on a few
         factors, namely: decoding and decompression. If the tuples are densely
         packed into data pages due to good encoding/compression ratios, there
         will be more work required when reconstructing the data. Each
         compression codec offers different performance tradeoffs and should be
         considered before writing the data. Due to the deterministic nature of
         the scheduler, single nodes can become bottlenecks for highly concurrent
         queries that use the same tables.
       </p>

       <p>
         If, for example, a Parquet based dataset is tiny, e.g. a small
         dimension table, such that it fits into a single HDFS block (Impala by
         default will create 256 MB blocks when Parquet is used, each containing
         a single row group) then there are a number of options that can be
         considered to resolve the potential scheduling hotspots when querying
         this data:
       </p>

       <ul>
         <li>
           In <keyword keyref="impala25"/> and higher, the scheduler’s
           deterministic behaviour can be changed using the following query
           options: <codeph>REPLICA_PREFERENCE</codeph> and
             <codeph>RANDOM_REPLICA</codeph>. For a detailed description of each
           of these modes see <keyword keyref="impala-2696">IMPALA-2696</keyword>.
         </li>

         <li>
           HDFS caching can be used to cache block replicas. This will cause
           the Impala scheduler to randomly pick (from <keyword keyref="impala22"
           /> and higher) a node that is hosting a cached block replica for the
           scan. Note, although HDFS caching has benefits, it serves only to help
           with the reading of raw block data and not cached tuple data, but with
           the right number of cached replicas (by default, HDFS only caches one
           replica), even load distribution can be achieved for smaller
           datasets.
         </li>

         <li>
           Do not compress the table data. The uncompressed table data spans more
           nodes and eliminates skew caused by compression.
         </li>

         <li>
           Reduce the Parquet file size via the
             <codeph>PARQUET_FILE_SIZE</codeph> query option when writing the
           table data. Using this approach the data will span more nodes. However
           it’s not recommended to drop the size below 32 MB.
         </li>
       </ul>
     </section>

   </conbody>
 </concept>
	<?xml version="1.0" encoding="UTF-8"?>
	<!--
	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.
	-->
	<!DOCTYPE concept PUBLIC "-//OASIS//DTD DITA Concept//EN" "concept.dtd">
	<concept id="perf_cookbook">

	<title>Impala Performance Guidelines and Best Practices</title>
	<titlealts audience="PDF"><navtitle>Performance Best Practices</navtitle></titlealts>
	<prolog>
	<metadata>
	<data name="Category" value="Performance"/>
	<data name="Category" value="Impala"/>
	<data name="Category" value="Planning"/>
	<data name="Category" value="Proof of Concept"/>
	<data name="Category" value="Guidelines"/>
	<data name="Category" value="Best Practices"/>
	<data name="Category" value="Proof of Concept"/>
	<data name="Category" value="Developers"/>
	<data name="Category" value="Data Analysts"/>
	</metadata>
	</prolog>

	<conbody>

	<p>
	Here are performance guidelines and best practices that you can use during planning, experimentation, and
	performance tuning for an Impala-enabled <keyword keyref="distro"/> cluster. All of this information is also available in more
	detail elsewhere in the Impala documentation; it is gathered together here to serve as a cookbook and
	emphasize which performance techniques typically provide the highest return on investment
	</p>

	<p outputclass="toc inpage"/>

	<section id="perf_cookbook_file_format">

	<title>Choose the appropriate file format for the data</title>

	<p>
	Typically, for large volumes of data (multiple gigabytes per table or partition), the Parquet file format
	performs best because of its combination of columnar storage layout, large I/O request size, and
	compression and encoding. See <xref href="impala_file_formats.xml#file_formats"/> for comparisons of all
	file formats supported by Impala, and <xref href="impala_parquet.xml#parquet"/> for details about the
	Parquet file format.
	</p>

	<note>
	For smaller volumes of data, a few gigabytes or less for each table or partition, you might not see
	significant performance differences between file formats. At small data volumes, reduced I/O from an
	efficient compressed file format can be counterbalanced by reduced opportunity for parallel execution. When
	planning for a production deployment or conducting benchmarks, always use realistic data volumes to get a
	true picture of performance and scalability.
	</note>
	</section>

	<section id="perf_cookbook_small_files">

	<title>Avoid data ingestion processes that produce many small files</title>

	<p>
	When producing data files outside of Impala, prefer either text format or Avro, where you can build up the
	files row by row. Once the data is in Impala, you can convert it to the more efficient Parquet format and
	split into multiple data files using a single <codeph>INSERT ... SELECT</codeph> statement. Or, if you have
	the infrastructure to produce multi-megabyte Parquet files as part of your data preparation process, do
	that and skip the conversion step inside Impala.
	</p>

	<p>
	Always use <codeph>INSERT ... SELECT</codeph> to copy significant volumes of data from table to table
	within Impala. Avoid <codeph>INSERT ... VALUES</codeph> for any substantial volume of data or
	performance-critical tables, because each such statement produces a separate tiny data file. See
	<xref href="impala_insert.xml#insert"/> for examples of the <codeph>INSERT ... SELECT</codeph> syntax.
	</p>

	<p>
	For example, if you have thousands of partitions in a Parquet table, each with less than
	<ph rev="parquet_block_size">256 MB</ph> of data, consider partitioning in a less granular way, such as by
	year / month rather than year / month / day. If an inefficient data ingestion process produces thousands of
	data files in the same table or partition, consider compacting the data by performing an <codeph>INSERT ...
	SELECT</codeph> to copy all the data to a different table; the data will be reorganized into a smaller
	number of larger files by this process.
	</p>
	</section>

	<section id="perf_cookbook_partitioning">

	<title>Choose partitioning granularity based on actual data volume</title>

	<p>
	Partitioning is a technique that physically divides the data based on values of one or more columns, such
	as by year, month, day, region, city, section of a web site, and so on. When you issue queries that request
	a specific value or range of values for the partition key columns, Impala can avoid reading the irrelevant
	data, potentially yielding a huge savings in disk I/O.
	</p>

	<p>
	When deciding which column(s) to use for partitioning, choose the right level of granularity. For example,
	should you partition by year, month, and day, or only by year and month? Choose a partitioning strategy
	that puts at least <ph rev="parquet_block_size">256 MB</ph> of data in each partition, to take advantage of
	HDFS bulk I/O and Impala distributed queries.
	</p>

	<p>
	Over-partitioning can also cause query planning to take longer than necessary, as Impala prunes the
	unnecessary partitions. Ideally, keep the number of partitions in the table under 30 thousand.
	</p>

	<p>
	When preparing data files to go in a partition directory, create several large files rather than many small
	ones. If you receive data in the form of many small files and have no control over the input format,
	consider using the <codeph>INSERT ... SELECT</codeph> syntax to copy data from one table or partition to
	another, which compacts the files into a relatively small number (based on the number of nodes in the
	cluster).
	</p>

	<p>
	If you need to reduce the overall number of partitions and increase the amount of data in each partition,
	first look for partition key columns that are rarely referenced or are referenced in non-critical queries
	(not subject to an SLA). For example, your web site log data might be partitioned by year, month, day, and
	hour, but if most queries roll up the results by day, perhaps you only need to partition by year, month,
	and day.
	</p>

	<p>
	If you need to reduce the granularity even more, consider creating <q>buckets</q>, computed values
	corresponding to different sets of partition key values. For example, you can use the
	<codeph>TRUNC()</codeph> function with a <codeph>TIMESTAMP</codeph> column to group date and time values
	based on intervals such as week or quarter. See
	<xref href="impala_datetime_functions.xml#datetime_functions"/> for details.
	</p>

	<p>
	See <xref href="impala_partitioning.xml#partitioning"/> for full details and performance considerations for
	partitioning.
	</p>
	</section>

	<section id="perf_cookbook_partition_keys">

	<title>Use smallest appropriate integer types for partition key columns</title>

	<p>
	Although it is tempting to use strings for partition key columns, since those values are turned into HDFS
	directory names anyway, you can minimize memory usage by using numeric values for common partition key
	fields such as <codeph>YEAR</codeph>, <codeph>MONTH</codeph>, and <codeph>DAY</codeph>. Use the smallest
	integer type that holds the appropriate range of values, typically <codeph>TINYINT</codeph> for
	<codeph>MONTH</codeph> and <codeph>DAY</codeph>, and <codeph>SMALLINT</codeph> for <codeph>YEAR</codeph>.
	Use the <codeph>EXTRACT()</codeph> function to pull out individual date and time fields from a
	<codeph>TIMESTAMP</codeph> value, and <codeph>CAST()</codeph> the return value to the appropriate integer
	type.
	</p>
	</section>

	<section id="perf_cookbook_parquet_block_size">

	<title>Choose an appropriate Parquet block size</title>

	<p rev="parquet_block_size">
	By default, the Impala <codeph>INSERT ... SELECT</codeph> statement creates Parquet files with a 256 MB
	block size. (This default was changed in Impala 2.0. Formerly, the limit was 1 GB, but Impala made
	conservative estimates about compression, resulting in files that were smaller than 1 GB.)
	</p>

	<p>
	Each Parquet file written by Impala is a single block, allowing the whole file to be processed as a unit by a single host.
	As you copy Parquet files into HDFS or between HDFS filesystems, use <codeph>hdfs dfs -pb</codeph> to preserve the original
	block size.
	</p>

	<p>
	If there is only one or a few data block in your Parquet table, or in a partition that is the only one
	accessed by a query, then you might experience a slowdown for a different reason: not enough data to take
	advantage of Impala's parallel distributed queries. Each data block is processed by a single core on one of
	the DataNodes. In a 100-node cluster of 16-core machines, you could potentially process thousands of data
	files simultaneously. You want to find a sweet spot between <q>many tiny files</q> and <q>single giant
	file</q> that balances bulk I/O and parallel processing. You can set the <codeph>PARQUET_FILE_SIZE</codeph>
	query option before doing an <codeph>INSERT ... SELECT</codeph> statement to reduce the size of each
	generated Parquet file. <ph rev="2.0.0">(Specify the file size as an absolute number of bytes, or in Impala
	2.0 and later, in units ending with <codeph>m</codeph> for megabytes or <codeph>g</codeph> for
	gigabytes.)</ph> Run benchmarks with different file sizes to find the right balance point for your
	particular data volume.
	</p>
	</section>

	<section id="perf_cookbook_stats">

	<title>Gather statistics for all tables used in performance-critical or high-volume join queries</title>

	<p>
	Gather the statistics with the <codeph>COMPUTE STATS</codeph> statement. See
	<xref href="impala_perf_joins.xml#perf_joins"/> for details.
	</p>
	</section>

	<section id="perf_cookbook_network">

	<title>Minimize the overhead of transmitting results back to the client</title>

	<p>
	Use techniques such as:
	</p>

	<ul>
	<li>
	Aggregation. If you need to know how many rows match a condition, the total values of matching values
	from some column, the lowest or highest matching value, and so on, call aggregate functions such as
	<codeph>COUNT()</codeph>, <codeph>SUM()</codeph>, and <codeph>MAX()</codeph> in the query rather than
	sending the result set to an application and doing those computations there. Remember that the size of an
	unaggregated result set could be huge, requiring substantial time to transmit across the network.
	</li>

	<li>
	Filtering. Use all applicable tests in the <codeph>WHERE</codeph> clause of a query to eliminate rows
	that are not relevant, rather than producing a big result set and filtering it using application logic.
	</li>

	<li>
	<codeph>LIMIT</codeph> clause. If you only need to see a few sample values from a result set, or the top
	or bottom values from a query using <codeph>ORDER BY</codeph>, include the <codeph>LIMIT</codeph> clause
	to reduce the size of the result set rather than asking for the full result set and then throwing most of
	the rows away.
	</li>

	<li>
	Avoid overhead from pretty-printing the result set and displaying it on the screen. When you retrieve the
	results through <cmdname>impala-shell</cmdname>, use <cmdname>impala-shell</cmdname> options such as
	<codeph>-B</codeph> and <codeph>--output_delimiter</codeph> to produce results without special
	formatting, and redirect output to a file rather than printing to the screen. Consider using
	<codeph>INSERT ... SELECT</codeph> to write the results directly to new files in HDFS. See
	<xref href="impala_shell_options.xml#shell_options"/> for details about the
	<cmdname>impala-shell</cmdname> command-line options.
	</li>
	</ul>
	</section>

	<section id="perf_cookbook_explain">

	<title>Verify that your queries are planned in an efficient logical manner</title>

	<p>
	Examine the <codeph>EXPLAIN</codeph> plan for a query before actually running it. See
	<xref href="impala_explain.xml#explain"/> and <xref href="impala_explain_plan.xml#perf_explain"/> for
	details.
	</p>
	</section>

	<section id="perf_cookbook_profile">

	<title>Verify performance characteristics of queries</title>

	<p>
	Verify that the low-level aspects of I/O, memory usage, network bandwidth, CPU utilization, and so on are
	within expected ranges by examining the query profile for a query after running it. See
	<xref href="impala_explain_plan.xml#perf_profile"/> for details.
	</p>
	</section>

	<section id="perf_cookbook_os">

	<title>Use appropriate operating system settings</title>

	<p>
	See <xref keyref="cdh_admin_performance"/> for recommendations about operating system
	settings that you can change to influence Impala performance. In particular, you might find
	that changing the <codeph>vm.swappiness</codeph> Linux kernel setting to a non-zero value improves
	overall performance.
	</p>
	</section>
	<section id="perf_cookbook_hotspot">
	<title>Hotspot analysis</title>

	<p>
	In the context of Impala, a hotspot is defined as “an Impala daemon
	that for a single query or a workload is spending a far greater amount
	of time processing data relative to its neighbours”.
	</p>

	<p>
	Before discussing the options to tackle this issue some background is
	first required to understand how this problem can occur.
	</p>

	<p>
	By default, the scheduling of scan based plan fragments is
	deterministic. This means that for multiple queries needing to read the
	same block of data, the same node will be picked to host the scan. The
	default scheduling logic does not take into account node workload from
	prior queries. The complexity of materializing a tuple depends on a few
	factors, namely: decoding and decompression. If the tuples are densely
	packed into data pages due to good encoding/compression ratios, there
	will be more work required when reconstructing the data. Each
	compression codec offers different performance tradeoffs and should be
	considered before writing the data. Due to the deterministic nature of
	the scheduler, single nodes can become bottlenecks for highly concurrent
	queries that use the same tables.
	</p>

	<p>
	If, for example, a Parquet based dataset is tiny, e.g. a small
	dimension table, such that it fits into a single HDFS block (Impala by
	default will create 256 MB blocks when Parquet is used, each containing
	a single row group) then there are a number of options that can be
	considered to resolve the potential scheduling hotspots when querying
	this data:
	</p>

	<ul>
	<li>
	In <keyword keyref="impala25"/> and higher, the scheduler’s
	deterministic behaviour can be changed using the following query
	options: <codeph>REPLICA_PREFERENCE</codeph> and
	<codeph>RANDOM_REPLICA</codeph>. For a detailed description of each
	of these modes see <keyword keyref="impala-2696">IMPALA-2696</keyword>.
	</li>

	<li>
	HDFS caching can be used to cache block replicas. This will cause
	the Impala scheduler to randomly pick (from <keyword keyref="impala22"
	/> and higher) a node that is hosting a cached block replica for the
	scan. Note, although HDFS caching has benefits, it serves only to help
	with the reading of raw block data and not cached tuple data, but with
	the right number of cached replicas (by default, HDFS only caches one
	replica), even load distribution can be achieved for smaller
	datasets.
	</li>

	<li>
	Do not compress the table data. The uncompressed table data spans more
	nodes and eliminates skew caused by compression.
	</li>

	<li>
	Reduce the Parquet file size via the
	<codeph>PARQUET_FILE_SIZE</codeph> query option when writing the
	table data. Using this approach the data will span more nodes. However
	it’s not recommended to drop the size below 32 MB.
	</li>
	</ul>
	</section>

	</conbody>
	</concept>