docs/topics/impala_default_join_distribution_mode.xml

<?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="default_join_distribution_mode" rev="2.9.0 IMPALA-5381 IMPALA-5583">

  <title>DEFAULT_JOIN_DISTRIBUTION_MODE Query Option</title>
  <titlealts audience="PDF"><navtitle>DEFAULT_JOIN_DISTRIBUTION_MODE</navtitle></titlealts>
  <prolog>
    <metadata>
      <data name="Category" value="Impala"/>
      <data name="Category" value="Impala Query Options"/>
      <data name="Category" value="Performance"/>
      <data name="Category" value="Querying"/>
      <data name="Category" value="Developers"/>
      <data name="Category" value="Data Analysts"/>
    </metadata>
  </prolog>

  <conbody>

    <p>
      <indexterm audience="hidden">DEFAULT_JOIN_DISTRIBUTION_MODE query option</indexterm>
      This option determines the join distribution that Impala uses when any of the tables
      involved in a join query is missing statistics.
    </p>

    <p>
      Impala optimizes join queries based on the presence of table statistics,
      which are produced by the Impala <codeph>COMPUTE STATS</codeph> statement.
      By default, when a table involved in the join query does not have statistics,
      Impala uses the <q>broadcast</q> technique that transmits the entire contents
      of the table to all executor nodes participating in the query. If one table
      involved in a join has statistics and the other does not, the table without
      statistics is broadcast. If both tables are missing statistics, the table
      that is referenced second in the join order is broadcast. This behavior
      is appropriate when the table involved is relatively small, but can lead to
      excessive network, memory, and CPU overhead if the table being broadcast is
      large.
    </p>

    <p>
      Because Impala queries frequently involve very large tables, and suboptimal
      joins for such tables could result in spilling or out-of-memory errors,
      the setting <codeph>DEFAULT_JOIN_DISTRIBUTION_MODE=SHUFFLE</codeph> lets you
      override the default behavior. The shuffle join mechanism divides the corresponding rows
      of each table involved in a join query using a hashing algorithm, and transmits
      subsets of the rows to other nodes for processing. Typically, this kind of join is
      more efficient for joins between large tables of similar size.
    </p>

    <p>
      The setting <codeph>DEFAULT_JOIN_DISTRIBUTION_MODE=SHUFFLE</codeph> is
      recommended when setting up and deploying new clusters, because it is less likely
      to result in serious consequences such as spilling or out-of-memory errors if
      the query plan is based on incomplete information. This setting is not the default,
      to avoid changing the performance characteristics of join queries for clusters that
      are already tuned for their existing workloads.
    </p>

    <p conref="../shared/impala_common.xml#common/type_integer"/>
    <p>
      The allowed values are <codeph>BROADCAST</codeph> (equivalent to 0)
      or <codeph>SHUFFLE</codeph> (equivalent to 1).
    </p>

    <p conref="../shared/impala_common.xml#common/example_blurb"/>
    <p>
      The following examples demonstrate appropriate scenarios for each
      setting of this query option.
    </p>

<codeblock>
-- Create a billion-row table.
create table big_table stored as parquet
  as select * from huge_table limit 1e9;

-- For a big table with no statistics, the
-- shuffle join mechanism is appropriate.
set default_join_distribution_mode=shuffle;

...join queries involving the big table...
</codeblock>

<codeblock>
-- Create a hundred-row table.
create table tiny_table stored as parquet
  as select * from huge_table limit 100;

-- For a tiny table with no statistics, the
-- broadcast join mechanism is appropriate.
set default_join_distribution_mode=broadcast;

...join queries involving the tiny table...
</codeblock>

<codeblock>
compute stats tiny_table;
compute stats big_table;

-- Once the stats are computed, the query option has
-- no effect on join queries involving these tables.
-- Impala can determine the absolute and relative sizes
-- of each side of the join query by examining the
-- row size, cardinality, and so on of each table.

...join queries involving both of these tables...
</codeblock>

    <p conref="../shared/impala_common.xml#common/related_info"/>
    <p>
      <xref keyref="compute_stats"/>,
      <xref keyref="joins"/>,
      <xref keyref="perf_joins"/>
    </p>

  </conbody>
</concept>