http://git-wip-us.apache.org/repos/asf/incubator-impala/blob/3be0f122/docs/topics/impala_perf_stats.xml ---------------------------------------------------------------------- diff --git a/docs/topics/impala_perf_stats.xml b/docs/topics/impala_perf_stats.xml new file mode 100644 index 0000000..33d394d --- /dev/null +++ b/docs/topics/impala_perf_stats.xml @@ -0,0 +1,1031 @@ +<?xml version="1.0" encoding="UTF-8"?> +<!DOCTYPE concept PUBLIC "-//OASIS//DTD DITA Concept//EN" "concept.dtd"> +<concept id="perf_stats"> + + <title>Table and Column Statistics</title> + <prolog> + <metadata> + <data name="Category" value="Impala"/> + <data name="Category" value="Performance"/> + <data name="Category" value="Querying"/> + <data name="Category" value="Concepts"/> + <data name="Category" value="Developers"/> + <data name="Category" value="Data Analysts"/> + </metadata> + </prolog> + + <conbody> + + <p> + Impala can do better optimization for complex or multi-table queries when it has access to statistics about + the volume of data and how the values are distributed. Impala uses this information to help parallelize and + distribute the work for a query. For example, optimizing join queries requires a way of determining if one + table is <q>bigger</q> than another, which is a function of the number of rows and the average row size + for each table. The following sections describe the categories of statistics Impala can work + with, and how to produce them and keep them up to date. + </p> + + <note> + <p rev="1.2.2"> + Originally, Impala relied on the Hive mechanism for collecting statistics, through the Hive <codeph>ANALYZE + TABLE</codeph> statement which initiates a MapReduce job. For better user-friendliness and reliability, + Impala implements its own <codeph>COMPUTE STATS</codeph> statement in Impala 1.2.2 and higher, along with the + <codeph>DROP STATS</codeph>, <codeph>SHOW TABLE STATS</codeph>, and <codeph>SHOW COLUMN STATS</codeph> + statements. + </p> + </note> + + <p outputclass="toc inpage"/> + </conbody> + + <concept id="perf_table_stats"> + + <title id="table_stats">Overview of Table Statistics</title> + <prolog> + <metadata> + <data name="Category" value="Concepts"/> + </metadata> + </prolog> + + <conbody> + +<!-- Hive background info: https://cwiki.apache.org/Hive/statsdev.html --> + + <p> + The Impala query planner can make use of statistics about entire tables and partitions. + This information includes physical characteristics such as the number of rows, number of data files, + the total size of the data files, and the file format. For partitioned tables, the numbers + are calculated per partition, and as totals for the whole table. + This metadata is stored in the metastore database, and can be updated by either Impala or Hive. + If a number is not available, the value -1 is used as a placeholder. + Some numbers, such as number and total sizes of data files, are always kept up to date because + they can be calculated cheaply, as part of gathering HDFS block metadata. + </p> + + <p> + The following example shows table stats for an unpartitioned Parquet table. + The values for the number and sizes of files are always available. + Initially, the number of rows is not known, because it requires a potentially expensive + scan through the entire table, and so that value is displayed as -1. + The <codeph>COMPUTE STATS</codeph> statement fills in any unknown table stats values. + </p> + +<codeblock> +show table stats parquet_snappy; ++-------+--------+---------+--------------+-------------------+---------+-------------------+... +| #Rows | #Files | Size | Bytes Cached | Cache Replication | Format | Incremental stats |... ++-------+--------+---------+--------------+-------------------+---------+-------------------+... +| -1 | 96 | 23.35GB | NOT CACHED | NOT CACHED | PARQUET | false |... ++-------+--------+---------+--------------+-------------------+---------+-------------------+... + +compute stats parquet_snappy; ++-----------------------------------------+ +| summary | ++-----------------------------------------+ +| Updated 1 partition(s) and 6 column(s). | ++-----------------------------------------+ + + +show table stats parquet_snappy; ++------------+--------+---------+--------------+-------------------+---------+-------------------+... +| #Rows | #Files | Size | Bytes Cached | Cache Replication | Format | Incremental stats |... ++------------+--------+---------+--------------+-------------------+---------+-------------------+... +| 1000000000 | 96 | 23.35GB | NOT CACHED | NOT CACHED | PARQUET | false |... ++------------+--------+---------+--------------+-------------------+---------+-------------------+... +</codeblock> + + <p> + Impala performs some optimizations using this metadata on its own, and other optimizations by + using a combination of table and column statistics. + </p> + + <p rev="1.2.1"> + To check that table statistics are available for a table, and see the details of those statistics, use the + statement <codeph>SHOW TABLE STATS <varname>table_name</varname></codeph>. See + <xref href="impala_show.xml#show"/> for details. + </p> + + <p> + If you use the Hive-based methods of gathering statistics, see + <xref href="https://cwiki.apache.org/confluence/display/Hive/StatsDev"; scope="external" format="html">the + Hive wiki</xref> for information about the required configuration on the Hive side. <ph rev="upstream">Cloudera</ph> recommends + using the Impala <codeph>COMPUTE STATS</codeph> statement to avoid potential configuration and scalability + issues with the statistics-gathering process. + </p> + + <p conref="../shared/impala_common.xml#common/hive_column_stats_caveat"/> + </conbody> + </concept> + + <concept id="perf_column_stats"> + + <title id="column_stats">Overview of Column Statistics</title> + + <conbody> + +<!-- Cloudera+Hive background information: http://blog.cloudera.com/blog/2012/08/column-statistics-in-hive/ --> + + <p> + The Impala query planner can make use of statistics about individual columns when that metadata is + available in the metastore database. This technique is most valuable for columns compared across tables in + <xref href="impala_perf_joins.xml#perf_joins">join queries</xref>, to help estimate how many rows the query + will retrieve from each table. <ph rev="2.0.0"> These statistics are also important for correlated + subqueries using the <codeph>EXISTS()</codeph> or <codeph>IN()</codeph> operators, which are processed + internally the same way as join queries.</ph> + </p> + + <p> + The following example shows column stats for an unpartitioned Parquet table. + The values for the maximum and average sizes of some types are always available, + because those figures are constant for numeric and other fixed-size types. + Initially, the number of distinct values is not known, because it requires a potentially expensive + scan through the entire table, and so that value is displayed as -1. + The same applies to maximum and average sizes of variable-sized types, such as <codeph>STRING</codeph>. + The <codeph>COMPUTE STATS</codeph> statement fills in most unknown column stats values. + (It does not record the number of <codeph>NULL</codeph> values, because currently Impala + does not use that figure for query optimization.) + </p> + +<codeblock> +show column stats parquet_snappy; ++-------------+----------+------------------+--------+----------+----------+ +| Column | Type | #Distinct Values | #Nulls | Max Size | Avg Size | ++-------------+----------+------------------+--------+----------+----------+ +| id | BIGINT | -1 | -1 | 8 | 8 | +| val | INT | -1 | -1 | 4 | 4 | +| zerofill | STRING | -1 | -1 | -1 | -1 | +| name | STRING | -1 | -1 | -1 | -1 | +| assertion | BOOLEAN | -1 | -1 | 1 | 1 | +| location_id | SMALLINT | -1 | -1 | 2 | 2 | ++-------------+----------+------------------+--------+----------+----------+ + +compute stats parquet_snappy; ++-----------------------------------------+ +| summary | ++-----------------------------------------+ +| Updated 1 partition(s) and 6 column(s). | ++-----------------------------------------+ + +show column stats parquet_snappy; ++-------------+----------+------------------+--------+----------+-------------------+ +| Column | Type | #Distinct Values | #Nulls | Max Size | Avg Size | ++-------------+----------+------------------+--------+----------+-------------------+ +| id | BIGINT | 183861280 | -1 | 8 | 8 | +| val | INT | 139017 | -1 | 4 | 4 | +| zerofill | STRING | 101761 | -1 | 6 | 6 | +| name | STRING | 145636240 | -1 | 22 | 13.00020027160645 | +| assertion | BOOLEAN | 2 | -1 | 1 | 1 | +| location_id | SMALLINT | 339 | -1 | 2 | 2 | ++-------------+----------+------------------+--------+----------+-------------------+ +</codeblock> + + <note> + <p> + For column statistics to be effective in Impala, you also need to have table statistics for the + applicable tables, as described in <xref href="impala_perf_stats.xml#perf_table_stats"/>. When you use + the Impala <codeph>COMPUTE STATS</codeph> statement, both table and column statistics are automatically + gathered at the same time, for all columns in the table. + </p> + <p conref="../shared/impala_common.xml#common/decimal_no_stats"/> + </note> + + <note conref="../shared/impala_common.xml#common/compute_stats_nulls"/> + +<!-- Hive-based instructions are considered obsolete since the introduction of the Impala COMPUTE STATS statement. + <p> + Add settings like the following to the <filepath>hive-site.xml</filepath> + configuration file, in the Hive configuration directory, on every node where you run + <codeph>ANALYZE TABLE</codeph> statements through the + <codeph>hive</codeph> shell. The + <codeph>hive.stats.ndv.error</codeph> setting represents the standard error when + estimating the number of distinct values for a column. The value of 5.0 is recommended as a tradeoff between the + accuracy of the gathered statistics and the resource usage of the stats-gathering process. + </p> + +<codeblock><![CDATA[<property> + <name>hive.stats.ndv.error</name> + <value>5.0</value> +</property>]]></codeblock> + + <p> + 5.0 is a relatively low value that devotes substantial computational resources to the statistics-gathering + process. To reduce the resource usage, you could increase this value; to make the statistics even more precise, + you could lower it. + </p> + + <p> + The syntax for gathering column statistics uses the <codeph>ANALYZE TABLE ... + COMPUTE STATISTICS</codeph> clause, with an additional <codeph>FOR + COLUMNS</codeph> clause. For partitioned tables, you can gather statistics for specific partitions by including + a clause <codeph>PARTITION + (<varname>col1=val1</varname>,<varname>col2=val2</varname>, + ...)</codeph>; but you cannot include the partitioning columns in the + <codeph>FOR COLUMNS</codeph> clause. Also, you cannot use fully qualified table + names, so issue a <codeph>USE</codeph> command first to switch to the + appropriate database. For example: + </p> + +<codeblock>USE <varname>database_name</varname>; +ANALYZE TABLE <varname>table_name</varname> COMPUTE STATISTICS FOR COLUMNS <varname>column_list</varname>; +ANALYZE TABLE <varname>table_name</varname> PARTITION (<varname>partition_specs</varname>) COMPUTE STATISTICS FOR COLUMNS <varname>column_list</varname>;</codeblock> +--> + + <p rev="1.2.1"> + To check whether column statistics are available for a particular set of columns, use the <codeph>SHOW + COLUMN STATS <varname>table_name</varname></codeph> statement, or check the extended + <codeph>EXPLAIN</codeph> output for a query against that table that refers to those columns. See + <xref href="impala_show.xml#show"/> and <xref href="impala_explain.xml#explain"/> for details. + </p> + + <p conref="../shared/impala_common.xml#common/hive_column_stats_caveat"/> + </conbody> + </concept> + + <concept id="perf_stats_partitions"> + <title id="stats_partitions">How Table and Column Statistics Work for Partitioned Tables</title> + <conbody> + + <p> + When you use Impala for <q>big data</q>, you are highly likely to use partitioning + for your biggest tables, the ones representing data that can be logically divided + based on dates, geographic regions, or similar criteria. The table and column statistics + are especially useful for optimizing queries on such tables. For example, a query involving + one year might involve substantially more or less data than a query involving a different year, + or a range of several years. Each query might be optimized differently as a result. + </p> + + <p> + The following examples show how table and column stats work with a partitioned table. + The table for this example is partitioned by year, month, and day. + For simplicity, the sample data consists of 5 partitions, all from the same year and month. + Table stats are collected independently for each partition. (In fact, the + <codeph>SHOW PARTITIONS</codeph> statement displays exactly the same information as + <codeph>SHOW TABLE STATS</codeph> for a partitioned table.) Column stats apply to + the entire table, not to individual partitions. Because the partition key column values + are represented as HDFS directories, their characteristics are typically known in advance, + even when the values for non-key columns are shown as -1. + </p> + +<codeblock> +show partitions year_month_day; ++-------+-------+-----+-------+--------+---------+--------------+-------------------+---------+... +| year | month | day | #Rows | #Files | Size | Bytes Cached | Cache Replication | Format |... ++-------+-------+-----+-------+--------+---------+--------------+-------------------+---------+... +| 2013 | 12 | 1 | -1 | 1 | 2.51MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 2 | -1 | 1 | 2.53MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 3 | -1 | 1 | 2.52MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 4 | -1 | 1 | 2.51MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 5 | -1 | 1 | 2.52MB | NOT CACHED | NOT CACHED | PARQUET |... +| Total | | | -1 | 5 | 12.58MB | 0B | | |... ++-------+-------+-----+-------+--------+---------+--------------+-------------------+---------+... + +show table stats year_month_day; ++-------+-------+-----+-------+--------+---------+--------------+-------------------+---------+... +| year | month | day | #Rows | #Files | Size | Bytes Cached | Cache Replication | Format |... ++-------+-------+-----+-------+--------+---------+--------------+-------------------+---------+... +| 2013 | 12 | 1 | -1 | 1 | 2.51MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 2 | -1 | 1 | 2.53MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 3 | -1 | 1 | 2.52MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 4 | -1 | 1 | 2.51MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 5 | -1 | 1 | 2.52MB | NOT CACHED | NOT CACHED | PARQUET |... +| Total | | | -1 | 5 | 12.58MB | 0B | | |... ++-------+-------+-----+-------+--------+---------+--------------+-------------------+---------+... + +show column stats year_month_day; ++-----------+---------+------------------+--------+----------+----------+ +| Column | Type | #Distinct Values | #Nulls | Max Size | Avg Size | ++-----------+---------+------------------+--------+----------+----------+ +| id | INT | -1 | -1 | 4 | 4 | +| val | INT | -1 | -1 | 4 | 4 | +| zfill | STRING | -1 | -1 | -1 | -1 | +| name | STRING | -1 | -1 | -1 | -1 | +| assertion | BOOLEAN | -1 | -1 | 1 | 1 | +| year | INT | 1 | 0 | 4 | 4 | +| month | INT | 1 | 0 | 4 | 4 | +| day | INT | 5 | 0 | 4 | 4 | ++-----------+---------+------------------+--------+----------+----------+ + +compute stats year_month_day; ++-----------------------------------------+ +| summary | ++-----------------------------------------+ +| Updated 5 partition(s) and 5 column(s). | ++-----------------------------------------+ + +show table stats year_month_day; ++-------+-------+-----+--------+--------+---------+--------------+-------------------+---------+... +| year | month | day | #Rows | #Files | Size | Bytes Cached | Cache Replication | Format |... ++-------+-------+-----+--------+--------+---------+--------------+-------------------+---------+... +| 2013 | 12 | 1 | 93606 | 1 | 2.51MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 2 | 94158 | 1 | 2.53MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 3 | 94122 | 1 | 2.52MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 4 | 93559 | 1 | 2.51MB | NOT CACHED | NOT CACHED | PARQUET |... +| 2013 | 12 | 5 | 93845 | 1 | 2.52MB | NOT CACHED | NOT CACHED | PARQUET |... +| Total | | | 469290 | 5 | 12.58MB | 0B | | |... ++-------+-------+-----+--------+--------+---------+--------------+-------------------+---------+... + +show column stats year_month_day; ++-----------+---------+------------------+--------+----------+-------------------+ +| Column | Type | #Distinct Values | #Nulls | Max Size | Avg Size | ++-----------+---------+------------------+--------+----------+-------------------+ +| id | INT | 511129 | -1 | 4 | 4 | +| val | INT | 364853 | -1 | 4 | 4 | +| zfill | STRING | 311430 | -1 | 6 | 6 | +| name | STRING | 471975 | -1 | 22 | 13.00160026550293 | +| assertion | BOOLEAN | 2 | -1 | 1 | 1 | +| year | INT | 1 | 0 | 4 | 4 | +| month | INT | 1 | 0 | 4 | 4 | +| day | INT | 5 | 0 | 4 | 4 | ++-----------+---------+------------------+--------+----------+-------------------+ +</codeblock> + + <note> + Partitioned tables can grow so large that scanning the entire table, as the <codeph>COMPUTE STATS</codeph> + statement does, is impractical just to update the statistics for a new partition. The standard + <codeph>COMPUTE STATS</codeph> statement might take hours, or even days. That situation is where you switch + to using incremental statistics, a feature available in <keyword keyref="impala21_full"/> and higher. + See <xref href="impala_perf_stats.xml#perf_stats_incremental"/> for details about this feature + and the <codeph>COMPUTE INCREMENTAL STATS</codeph> syntax. + </note> + + <p conref="../shared/impala_common.xml#common/hive_column_stats_caveat"/> + </conbody> + </concept> + + <concept rev="2.1.0" id="perf_stats_incremental"> + + <title id="incremental_stats">Overview of Incremental Statistics</title> + + <conbody> + + <p> + In Impala 2.1.0 and higher, you can use the syntax <codeph>COMPUTE INCREMENTAL STATS</codeph> and + <codeph>DROP INCREMENTAL STATS</codeph>. The <codeph>INCREMENTAL</codeph> clauses work with incremental + statistics, a specialized feature for partitioned tables that are large or frequently updated with new + partitions. + </p> + + <p> + When you compute incremental statistics for a partitioned table, by default Impala only processes those + partitions that do not yet have incremental statistics. By processing only newly added partitions, you can + keep statistics up to date for large partitioned tables, without incurring the overhead of reprocessing the + entire table each time. + </p> + + <p> + You can also compute or drop statistics for a single partition by including a <codeph>PARTITION</codeph> + clause in the <codeph>COMPUTE INCREMENTAL STATS</codeph> or <codeph>DROP INCREMENTAL STATS</codeph> + statement. + </p> + + <p> + The metadata for incremental statistics is handled differently from the original style of statistics: + </p> + + <ul> + <li> + <p> + If you have an existing partitioned table for which you have already computed statistics, issuing + <codeph>COMPUTE INCREMENTAL STATS</codeph> without a partition clause causes Impala to rescan the + entire table. Once the incremental statistics are computed, any future <codeph>COMPUTE INCREMENTAL + STATS</codeph> statements only scan any new partitions and any partitions where you performed + <codeph>DROP INCREMENTAL STATS</codeph>. + </p> + </li> + + <li> + <p> + The <codeph>SHOW TABLE STATS</codeph> and <codeph>SHOW PARTITIONS</codeph> statements now include an + additional column showing whether incremental statistics are available for each column. A partition + could already be covered by the original type of statistics based on a prior <codeph>COMPUTE + STATS</codeph> statement, as indicated by a value other than <codeph>-1</codeph> under the + <codeph>#Rows</codeph> column. Impala query planning uses either kind of statistics when available. + </p> + </li> + + <li> + <p> + <codeph>COMPUTE INCREMENTAL STATS</codeph> takes more time than <codeph>COMPUTE STATS</codeph> for the + same volume of data. Therefore it is most suitable for tables with large data volume where new + partitions are added frequently, making it impractical to run a full <codeph>COMPUTE STATS</codeph> + operation for each new partition. For unpartitioned tables, or partitioned tables that are loaded once + and not updated with new partitions, use the original <codeph>COMPUTE STATS</codeph> syntax. + </p> + </li> + + <li> + <p> + <codeph>COMPUTE INCREMENTAL STATS</codeph> uses some memory in the <cmdname>catalogd</cmdname> process, + proportional to the number of partitions and number of columns in the applicable table. The memory + overhead is approximately 400 bytes for each column in each partition. This memory is reserved in the + <cmdname>catalogd</cmdname> daemon, the <cmdname>statestored</cmdname> daemon, and in each instance of + the <cmdname>impalad</cmdname> daemon. + </p> + </li> + + <li> + <p> + In cases where new files are added to an existing partition, issue a <codeph>REFRESH</codeph> statement + for the table, followed by a <codeph>DROP INCREMENTAL STATS</codeph> and <codeph>COMPUTE INCREMENTAL + STATS</codeph> sequence for the changed partition. + </p> + </li> + + <li> + <p> + The <codeph>DROP INCREMENTAL STATS</codeph> statement operates only on a single partition at a time. To + remove statistics (whether incremental or not) from all partitions of a table, issue a <codeph>DROP + STATS</codeph> statement with no <codeph>INCREMENTAL</codeph> or <codeph>PARTITION</codeph> clauses. + </p> + </li> + </ul> + + <p> + The following considerations apply to incremental statistics when the structure of an existing table is + changed (known as <term>schema evolution</term>): + </p> + + <ul> + <li> + <p> + If you use an <codeph>ALTER TABLE</codeph> statement to drop a column, the existing statistics remain + valid and <codeph>COMPUTE INCREMENTAL STATS</codeph> does not rescan any partitions. + </p> + </li> + + <li> + <p> + If you use an <codeph>ALTER TABLE</codeph> statement to add a column, Impala rescans all partitions and + fills in the appropriate column-level values the next time you run <codeph>COMPUTE INCREMENTAL + STATS</codeph>. + </p> + </li> + + <li> + <p> + If you use an <codeph>ALTER TABLE</codeph> statement to change the data type of a column, Impala + rescans all partitions and fills in the appropriate column-level values the next time you run + <codeph>COMPUTE INCREMENTAL STATS</codeph>. + </p> + </li> + + <li> + <p> + If you use an <codeph>ALTER TABLE</codeph> statement to change the file format of a table, the existing + statistics remain valid and a subsequent <codeph>COMPUTE INCREMENTAL STATS</codeph> does not rescan any + partitions. + </p> + </li> + </ul> + + <p> + See <xref href="impala_compute_stats.xml#compute_stats"/> and + <xref href="impala_drop_stats.xml#drop_stats"/> for syntax details. + </p> + </conbody> + </concept> + + <concept id="perf_stats_computing"> + <title>Generating Table and Column Statistics (COMPUTE STATS Statement)</title> + <conbody> + + <p> + To gather table statistics after loading data into a table or partition, you typically use the + <codeph>COMPUTE STATS</codeph> statement. This statement is available in Impala 1.2.2 and higher. + It gathers both table statistics and column statistics for all columns in a single operation. + For large partitioned tables, where you frequently need to update statistics and it is impractical + to scan the entire table each time, use the syntax <codeph>COMPUTE INCREMENTAL STATS</codeph>, + which is available in <keyword keyref="impala21_full"/> and higher. + </p> + + <p> + If you use Hive as part of your ETL workflow, you can also use Hive to generate table and + column statistics. You might need to do extra configuration within Hive itself, the metastore, + or even set up a separate database to hold Hive-generated statistics. You might need to run + multiple statements to generate all the necessary statistics. Therefore, prefer the + Impala <codeph>COMPUTE STATS</codeph> statement where that technique is practical. + For details about collecting statistics through Hive, see + <xref href="https://cwiki.apache.org/confluence/display/Hive/StatsDev"; scope="external" format="html">the Hive wiki</xref>. + </p> + + <p conref="../shared/impala_common.xml#common/hive_column_stats_caveat"/> + +<!-- Commenting out over-detailed Hive instructions as part of stats reorg. + <li> + Issue an <codeph>ANALYZE TABLE</codeph> statement in Hive, for the entire table or a specific partition. +<codeblock>ANALYZE TABLE <varname>tablename</varname> [PARTITION(<varname>partcol1</varname>[=<varname>val1</varname>], <varname>partcol2</varname>[=<varname>val2</varname>], ...)] COMPUTE STATISTICS [NOSCAN];</codeblock> + For example, to gather statistics for a non-partitioned table: +<codeblock>ANALYZE TABLE customer COMPUTE STATISTICS;</codeblock> + To gather statistics for a <codeph>store</codeph> table partitioned by state and city, and both of its + partitions: +<codeblock>ANALYZE TABLE store PARTITION(s_state, s_county) COMPUTE STATISTICS;</codeblock> + To gather statistics for the <codeph>store</codeph> table and only the partitions for California: +<codeblock>ANALYZE TABLE store PARTITION(s_state='CA', s_county) COMPUTE STATISTICS;</codeblock> + </li> + + <li> + Load the data through the <codeph>INSERT OVERWRITE</codeph> statement in Hive, while the Hive setting + <b>hive.stats.autogather</b> is enabled. + </li> + + </ul> +--> + + <p rev="2.0.1"> +<!-- Additional info as a result of IMPALA-1420 --> +<!-- Keep checking if https://issues.apache.org/jira/browse/HIVE-8648 ever gets fixed and when that fix makes it into a CDH release. --> + For your very largest tables, you might find that <codeph>COMPUTE STATS</codeph> or even <codeph>COMPUTE INCREMENTAL STATS</codeph> + take so long to scan the data that it is impractical to use them regularly. In such a case, after adding a partition or inserting new data, + you can update just the number of rows property through an <codeph>ALTER TABLE</codeph> statement. + See <xref href="impala_perf_stats.xml#perf_table_stats_manual"/> for details. + Because the column statistics might be left in a stale state, do not use this technique as a replacement + for <codeph>COMPUTE STATS</codeph>. Only use this technique if all other means of collecting statistics are impractical, or as a + low-overhead operation that you run in between periodic <codeph>COMPUTE STATS</codeph> or <codeph>COMPUTE INCREMENTAL STATS</codeph> operations. + </p> + + </conbody> + </concept> + + <concept rev="2.1.0" id="perf_stats_checking"> + + <title>Detecting Missing Statistics</title> + + <conbody> + + <p> + You can check whether a specific table has statistics using the <codeph>SHOW TABLE STATS</codeph> statement + (for any table) or the <codeph>SHOW PARTITIONS</codeph> statement (for a partitioned table). Both + statements display the same information. If a table or a partition does not have any statistics, the + <codeph>#Rows</codeph> field contains <codeph>-1</codeph>. Once you compute statistics for the table or + partition, the <codeph>#Rows</codeph> field changes to an accurate value. + </p> + + <p> + The following example shows a table that initially does not have any statistics. The <codeph>SHOW TABLE + STATS</codeph> statement displays different values for <codeph>#Rows</codeph> before and after the + <codeph>COMPUTE STATS</codeph> operation. + </p> + +<codeblock>[localhost:21000] > create table no_stats (x int); +[localhost:21000] > show table stats no_stats; ++-------+--------+------+--------------+--------+-------------------+ +| #Rows | #Files | Size | Bytes Cached | Format | Incremental stats | ++-------+--------+------+--------------+--------+-------------------+ +| -1 | 0 | 0B | NOT CACHED | TEXT | false | ++-------+--------+------+--------------+--------+-------------------+ +[localhost:21000] > compute stats no_stats; ++-----------------------------------------+ +| summary | ++-----------------------------------------+ +| Updated 1 partition(s) and 1 column(s). | ++-----------------------------------------+ +[localhost:21000] > show table stats no_stats; ++-------+--------+------+--------------+--------+-------------------+ +| #Rows | #Files | Size | Bytes Cached | Format | Incremental stats | ++-------+--------+------+--------------+--------+-------------------+ +| 0 | 0 | 0B | NOT CACHED | TEXT | false | ++-------+--------+------+--------------+--------+-------------------+ +</codeblock> + + <p> + The following example shows a similar progression with a partitioned table. Initially, + <codeph>#Rows</codeph> is <codeph>-1</codeph>. After a <codeph>COMPUTE STATS</codeph> operation, + <codeph>#Rows</codeph> changes to an accurate value. Any newly added partition starts with no statistics, + meaning that you must collect statistics after adding a new partition. + </p> + +<codeblock>[localhost:21000] > create table no_stats_partitioned (x int) partitioned by (year smallint); +[localhost:21000] > show table stats no_stats_partitioned; ++-------+-------+--------+------+--------------+--------+-------------------+ +| year | #Rows | #Files | Size | Bytes Cached | Format | Incremental stats | ++-------+-------+--------+------+--------------+--------+-------------------+ +| Total | -1 | 0 | 0B | 0B | | | ++-------+-------+--------+------+--------------+--------+-------------------+ +[localhost:21000] > show partitions no_stats_partitioned; ++-------+-------+--------+------+--------------+--------+-------------------+ +| year | #Rows | #Files | Size | Bytes Cached | Format | Incremental stats | ++-------+-------+--------+------+--------------+--------+-------------------+ +| Total | -1 | 0 | 0B | 0B | | | ++-------+-------+--------+------+--------------+--------+-------------------+ +[localhost:21000] > alter table no_stats_partitioned add partition (year=2013); +[localhost:21000] > compute stats no_stats_partitioned; ++-----------------------------------------+ +| summary | ++-----------------------------------------+ +| Updated 1 partition(s) and 1 column(s). | ++-----------------------------------------+ +[localhost:21000] > alter table no_stats_partitioned add partition (year=2014); +[localhost:21000] > show partitions no_stats_partitioned; ++-------+-------+--------+------+--------------+--------+-------------------+ +| year | #Rows | #Files | Size | Bytes Cached | Format | Incremental stats | ++-------+-------+--------+------+--------------+--------+-------------------+ +| 2013 | 0 | 0 | 0B | NOT CACHED | TEXT | false | +| 2014 | -1 | 0 | 0B | NOT CACHED | TEXT | false | +| Total | 0 | 0 | 0B | 0B | | | ++-------+-------+--------+------+--------------+--------+-------------------+ +</codeblock> + + <note> + Because the default <codeph>COMPUTE STATS</codeph> statement creates and updates statistics for all + partitions in a table, if you expect to frequently add new partitions, use the <codeph>COMPUTE INCREMENTAL + STATS</codeph> syntax instead, which lets you compute stats for a single specified partition, or only for + those partitions that do not already have incremental stats. + </note> + + <p> + If checking each individual table is impractical, due to a large number of tables or views that hide the + underlying base tables, you can also check for missing statistics for a particular query. Use the + <codeph>EXPLAIN</codeph> statement to preview query efficiency before actually running the query. Use the + query profile output available through the <codeph>PROFILE</codeph> command in + <cmdname>impala-shell</cmdname> or the web UI to verify query execution and timing after running the query. + Both the <codeph>EXPLAIN</codeph> plan and the <codeph>PROFILE</codeph> output display a warning if any + tables or partitions involved in the query do not have statistics. + </p> + +<codeblock>[localhost:21000] > create table no_stats (x int); +[localhost:21000] > explain select count(*) from no_stats; ++------------------------------------------------------------------------------------+ +| Explain String | ++------------------------------------------------------------------------------------+ +| Estimated Per-Host Requirements: Memory=10.00MB VCores=1 | +| WARNING: The following tables are missing relevant table and/or column statistics. | +| incremental_stats.no_stats | +| | +| 03:AGGREGATE [FINALIZE] | +| | output: count:merge(*) | +| | | +| 02:EXCHANGE [UNPARTITIONED] | +| | | +| 01:AGGREGATE | +| | output: count(*) | +| | | +| 00:SCAN HDFS [incremental_stats.no_stats] | +| partitions=1/1 files=0 size=0B | ++------------------------------------------------------------------------------------+ +</codeblock> + + <p> + Because Impala uses the <term>partition pruning</term> technique when possible to only evaluate certain + partitions, if you have a partitioned table with statistics for some partitions and not others, whether or + not the <codeph>EXPLAIN</codeph> statement shows the warning depends on the actual partitions used by the + query. For example, you might see warnings or not for different queries against the same table: + </p> + +<codeblock>-- No warning because all the partitions for the year 2012 have stats. +EXPLAIN SELECT ... FROM t1 WHERE year = 2012; + +-- Missing stats warning because one or more partitions in this range +-- do not have stats. +EXPLAIN SELECT ... FROM t1 WHERE year BETWEEN 2006 AND 2009; +</codeblock> + + <p> + To confirm if any partitions at all in the table are missing statistics, you might explain a query that + scans the entire table, such as <codeph>SELECT COUNT(*) FROM <varname>table_name</varname></codeph>. + </p> + </conbody> + </concept> + + <concept rev="2.1.0" id="perf_stats_collecting"> + + <title>Keeping Statistics Up to Date</title> + + <conbody> + + <p> + When the contents of a table or partition change significantly, recompute the stats for the relevant table + or partition. The degree of change that qualifies as <q>significant</q> varies, depending on the absolute + and relative sizes of the tables. Typically, if you add more than 30% more data to a table, it is + worthwhile to recompute stats, because the differences in number of rows and number of distinct values + might cause Impala to choose a different join order when that table is used in join queries. This guideline + is most important for the largest tables. For example, adding 30% new data to a table containing 1 TB has a + greater effect on join order than adding 30% to a table containing only a few megabytes, and the larger + table has a greater effect on query performance if Impala chooses a suboptimal join order as a result of + outdated statistics. + </p> + + <p> + If you reload a complete new set of data for a table, but the number of rows and number of distinct values + for each column is relatively unchanged from before, you do not need to recompute stats for the table. + </p> + + <p> + If the statistics for a table are out of date, and the table's large size makes it impractical to recompute + new stats immediately, you can use the <codeph>DROP STATS</codeph> statement to remove the obsolete + statistics, making it easier to identify tables that need a new <codeph>COMPUTE STATS</codeph> operation. + </p> + + <p> + For a large partitioned table, consider using the incremental stats feature available in Impala 2.1.0 and + higher, as explained in <xref href="impala_perf_stats.xml#perf_stats_incremental"/>. If you add a new + partition to a table, it is worthwhile to recompute incremental stats, because the operation only scans the + data for that one new partition. + </p> + </conbody> + </concept> + +<!-- Might deserve its own conceptual topic at some point. --> + + <concept audience="Cloudera" rev="1.2.2" id="perf_stats_joins"> + + <title>How Statistics Are Used in Join Queries</title> + + <conbody> + + <p></p> + </conbody> + </concept> + +<!-- Might deserve its own conceptual topic at some point. --> + + <concept audience="Cloudera" rev="1.2.2" id="perf_stats_inserts"> + + <title>How Statistics Are Used in INSERT Operations</title> + + <conbody> + + <p conref="../shared/impala_common.xml#common/insert_hints"/> + </conbody> + </concept> + + <concept rev="1.2.2" id="perf_table_stats_manual"> + + <title>Setting the NUMROWS Value Manually through ALTER TABLE</title> + + <conbody> + + <p> + The most crucial piece of data in all the statistics is the number of rows in the table (for an + unpartitioned or partitioned table) and for each partition (for a partitioned table). The <codeph>COMPUTE STATS</codeph> + statement always gathers statistics about all columns, as well as overall table statistics. If it is not + practical to do a full <codeph>COMPUTE STATS</codeph> or <codeph>COMPUTE INCREMENTAL STATS</codeph> + operation after adding a partition or inserting data, or if you can see that Impala would produce a more + efficient plan if the number of rows was different, you can manually set the number of rows through an + <codeph>ALTER TABLE</codeph> statement: + </p> + +<codeblock> +-- Set total number of rows. Applies to both unpartitioned and partitioned tables. +alter table <varname>table_name</varname> set tblproperties('numRows'='<varname>new_value</varname>', 'STATS_GENERATED_VIA_STATS_TASK'='true'); + +-- Set total number of rows for a specific partition. Applies to partitioned tables only. +-- You must specify all the partition key columns in the PARTITION clause. +alter table <varname>table_name</varname> partition (<varname>keycol1</varname>=<varname>val1</varname>,<varname>keycol2</varname>=<varname>val2</varname>...) set tblproperties('numRows'='<varname>new_value</varname>', 'STATS_GENERATED_VIA_STATS_TASK'='true'); +</codeblock> + + <p> + This statement avoids re-scanning any data files. (The requirement to include the <codeph>STATS_GENERATED_VIA_STATS_TASK</codeph> property is relatively new, as a + result of the issue <xref href="https://issues.apache.org/jira/browse/HIVE-8648"; scope="external" format="html">HIVE-8648</xref> + for the Hive metastore.) + </p> + +<codeblock conref="../shared/impala_common.xml#common/set_numrows_example"/> + + <p> + For a partitioned table, update both the per-partition number of rows and the number of rows for the whole + table: + </p> + +<codeblock conref="../shared/impala_common.xml#common/set_numrows_partitioned_example"/> + + <p> + In practice, the <codeph>COMPUTE STATS</codeph> statement, or <codeph>COMPUTE INCREMENTAL STATS</codeph> + for a partitioned table, should be fast and convenient enough that this technique is only useful for the very + largest partitioned tables. + <!-- + It is most useful as a workaround for in case of performance issues where you might adjust the <codeph>numRows</codeph> value higher + or lower to produce the ideal join order. + --> + <!-- Following wording is duplicated from earlier. Consider conref'ing. --> + Because the column statistics might be left in a stale state, do not use this technique as a replacement + for <codeph>COMPUTE STATS</codeph>. Only use this technique if all other means of collecting statistics are impractical, or as a + low-overhead operation that you run in between periodic <codeph>COMPUTE STATS</codeph> or <codeph>COMPUTE INCREMENTAL STATS</codeph> operations. + </p> + </conbody> + </concept> + + <concept id="perf_column_stats_manual" rev="2.6.0 IMPALA-3369"> + <title>Setting Column Stats Manually through ALTER TABLE</title> + <conbody> + <p> + In <keyword keyref="impala26_full"/> and higher, you can also use the <codeph>SET COLUMN STATS</codeph> + clause of <codeph>ALTER TABLE</codeph> to manually set or change column statistics. + Only use this technique in cases where it is impractical to run + <codeph>COMPUTE STATS</codeph> or <codeph>COMPUTE INCREMENTAL STATS</codeph> + frequently enough to keep up with data changes for a huge table. + </p> + <p conref="../shared/impala_common.xml#common/set_column_stats_example"/> + </conbody> + </concept> + + <concept rev="1.2.2" id="perf_stats_examples"> + + <title>Examples of Using Table and Column Statistics with Impala</title> + + <conbody> + + <p> + The following examples walk through a sequence of <codeph>SHOW TABLE STATS</codeph>, <codeph>SHOW COLUMN + STATS</codeph>, <codeph>ALTER TABLE</codeph>, and <codeph>SELECT</codeph> and <codeph>INSERT</codeph> + statements to illustrate various aspects of how Impala uses statistics to help optimize queries. + </p> + + <p> + This example shows table and column statistics for the <codeph>STORE</codeph> column used in the + <xref href="http://www.tpc.org/tpcds/"; scope="external" format="html">TPC-DS benchmarks for decision + support</xref> systems. It is a tiny table holding data for 12 stores. Initially, before any statistics are + gathered by a <codeph>COMPUTE STATS</codeph> statement, most of the numeric fields show placeholder values + of -1, indicating that the figures are unknown. The figures that are filled in are values that are easily + countable or deducible at the physical level, such as the number of files, total data size of the files, + and the maximum and average sizes for data types that have a constant size such as <codeph>INT</codeph>, + <codeph>FLOAT</codeph>, and <codeph>TIMESTAMP</codeph>. + </p> + +<codeblock>[localhost:21000] > show table stats store; ++-------+--------+--------+--------+ +| #Rows | #Files | Size | Format | ++-------+--------+--------+--------+ +| -1 | 1 | 3.08KB | TEXT | ++-------+--------+--------+--------+ +Returned 1 row(s) in 0.03s +[localhost:21000] > show column stats store; ++--------------------+-----------+------------------+--------+----------+----------+ +| Column | Type | #Distinct Values | #Nulls | Max Size | Avg Size | ++--------------------+-----------+------------------+--------+----------+----------+ +| s_store_sk | INT | -1 | -1 | 4 | 4 | +| s_store_id | STRING | -1 | -1 | -1 | -1 | +| s_rec_start_date | TIMESTAMP | -1 | -1 | 16 | 16 | +| s_rec_end_date | TIMESTAMP | -1 | -1 | 16 | 16 | +| s_closed_date_sk | INT | -1 | -1 | 4 | 4 | +| s_store_name | STRING | -1 | -1 | -1 | -1 | +| s_number_employees | INT | -1 | -1 | 4 | 4 | +| s_floor_space | INT | -1 | -1 | 4 | 4 | +| s_hours | STRING | -1 | -1 | -1 | -1 | +| s_manager | STRING | -1 | -1 | -1 | -1 | +| s_market_id | INT | -1 | -1 | 4 | 4 | +| s_geography_class | STRING | -1 | -1 | -1 | -1 | +| s_market_desc | STRING | -1 | -1 | -1 | -1 | +| s_market_manager | STRING | -1 | -1 | -1 | -1 | +| s_division_id | INT | -1 | -1 | 4 | 4 | +| s_division_name | STRING | -1 | -1 | -1 | -1 | +| s_company_id | INT | -1 | -1 | 4 | 4 | +| s_company_name | STRING | -1 | -1 | -1 | -1 | +| s_street_number | STRING | -1 | -1 | -1 | -1 | +| s_street_name | STRING | -1 | -1 | -1 | -1 | +| s_street_type | STRING | -1 | -1 | -1 | -1 | +| s_suite_number | STRING | -1 | -1 | -1 | -1 | +| s_city | STRING | -1 | -1 | -1 | -1 | +| s_county | STRING | -1 | -1 | -1 | -1 | +| s_state | STRING | -1 | -1 | -1 | -1 | +| s_zip | STRING | -1 | -1 | -1 | -1 | +| s_country | STRING | -1 | -1 | -1 | -1 | +| s_gmt_offset | FLOAT | -1 | -1 | 4 | 4 | +| s_tax_percentage | FLOAT | -1 | -1 | 4 | 4 | ++--------------------+-----------+------------------+--------+----------+----------+ +Returned 29 row(s) in 0.04s</codeblock> + + <p> + With the Hive <codeph>ANALYZE TABLE</codeph> statement for column statistics, you had to specify each + column for which to gather statistics. The Impala <codeph>COMPUTE STATS</codeph> statement automatically + gathers statistics for all columns, because it reads through the entire table relatively quickly and can + efficiently compute the values for all the columns. This example shows how after running the + <codeph>COMPUTE STATS</codeph> statement, statistics are filled in for both the table and all its columns: + </p> + +<codeblock>[localhost:21000] > compute stats store; ++------------------------------------------+ +| summary | ++------------------------------------------+ +| Updated 1 partition(s) and 29 column(s). | ++------------------------------------------+ +Returned 1 row(s) in 1.88s +[localhost:21000] > show table stats store; ++-------+--------+--------+--------+ +| #Rows | #Files | Size | Format | ++-------+--------+--------+--------+ +| 12 | 1 | 3.08KB | TEXT | ++-------+--------+--------+--------+ +Returned 1 row(s) in 0.02s +[localhost:21000] > show column stats store; ++--------------------+-----------+------------------+--------+----------+-------------------+ +| Column | Type | #Distinct Values | #Nulls | Max Size | Avg Size | ++--------------------+-----------+------------------+--------+----------+-------------------+ +| s_store_sk | INT | 12 | -1 | 4 | 4 | +| s_store_id | STRING | 6 | -1 | 16 | 16 | +| s_rec_start_date | TIMESTAMP | 4 | -1 | 16 | 16 | +| s_rec_end_date | TIMESTAMP | 3 | -1 | 16 | 16 | +| s_closed_date_sk | INT | 3 | -1 | 4 | 4 | +| s_store_name | STRING | 8 | -1 | 5 | 4.25 | +| s_number_employees | INT | 9 | -1 | 4 | 4 | +| s_floor_space | INT | 10 | -1 | 4 | 4 | +| s_hours | STRING | 2 | -1 | 8 | 7.083300113677979 | +| s_manager | STRING | 7 | -1 | 15 | 12 | +| s_market_id | INT | 7 | -1 | 4 | 4 | +| s_geography_class | STRING | 1 | -1 | 7 | 7 | +| s_market_desc | STRING | 10 | -1 | 94 | 55.5 | +| s_market_manager | STRING | 7 | -1 | 16 | 14 | +| s_division_id | INT | 1 | -1 | 4 | 4 | +| s_division_name | STRING | 1 | -1 | 7 | 7 | +| s_company_id | INT | 1 | -1 | 4 | 4 | +| s_company_name | STRING | 1 | -1 | 7 | 7 | +| s_street_number | STRING | 9 | -1 | 3 | 2.833300113677979 | +| s_street_name | STRING | 12 | -1 | 11 | 6.583300113677979 | +| s_street_type | STRING | 8 | -1 | 9 | 4.833300113677979 | +| s_suite_number | STRING | 11 | -1 | 9 | 8.25 | +| s_city | STRING | 2 | -1 | 8 | 6.5 | +| s_county | STRING | 1 | -1 | 17 | 17 | +| s_state | STRING | 1 | -1 | 2 | 2 | +| s_zip | STRING | 2 | -1 | 5 | 5 | +| s_country | STRING | 1 | -1 | 13 | 13 | +| s_gmt_offset | FLOAT | 1 | -1 | 4 | 4 | +| s_tax_percentage | FLOAT | 5 | -1 | 4 | 4 | ++--------------------+-----------+------------------+--------+----------+-------------------+ +Returned 29 row(s) in 0.04s</codeblock> + + <p> + The following example shows how statistics are represented for a partitioned table. In this case, we have + set up a table to hold the world's most trivial census data, a single <codeph>STRING</codeph> field, + partitioned by a <codeph>YEAR</codeph> column. The table statistics include a separate entry for each + partition, plus final totals for the numeric fields. The column statistics include some easily deducible + facts for the partitioning column, such as the number of distinct values (the number of partition + subdirectories). +<!-- and the number of <codeph>NULL</codeph> values (none in this case). --> + </p> + +<codeblock>localhost:21000] > describe census; ++------+----------+---------+ +| name | type | comment | ++------+----------+---------+ +| name | string | | +| year | smallint | | ++------+----------+---------+ +Returned 2 row(s) in 0.02s +[localhost:21000] > show table stats census; ++-------+-------+--------+------+---------+ +| year | #Rows | #Files | Size | Format | ++-------+-------+--------+------+---------+ +| 2000 | -1 | 0 | 0B | TEXT | +| 2004 | -1 | 0 | 0B | TEXT | +| 2008 | -1 | 0 | 0B | TEXT | +| 2010 | -1 | 0 | 0B | TEXT | +| 2011 | 0 | 1 | 22B | TEXT | +| 2012 | -1 | 1 | 22B | TEXT | +| 2013 | -1 | 1 | 231B | PARQUET | +| Total | 0 | 3 | 275B | | ++-------+-------+--------+------+---------+ +Returned 8 row(s) in 0.02s +[localhost:21000] > show column stats census; ++--------+----------+------------------+--------+----------+----------+ +| Column | Type | #Distinct Values | #Nulls | Max Size | Avg Size | ++--------+----------+------------------+--------+----------+----------+ +| name | STRING | -1 | -1 | -1 | -1 | +| year | SMALLINT | 7 | -1 | 2 | 2 | ++--------+----------+------------------+--------+----------+----------+ +Returned 2 row(s) in 0.02s</codeblock> + + <p> + The following example shows how the statistics are filled in by a <codeph>COMPUTE STATS</codeph> statement + in Impala. + </p> + +<codeblock>[localhost:21000] > compute stats census; ++-----------------------------------------+ +| summary | ++-----------------------------------------+ +| Updated 3 partition(s) and 1 column(s). | ++-----------------------------------------+ +Returned 1 row(s) in 2.16s +[localhost:21000] > show table stats census; ++-------+-------+--------+------+---------+ +| year | #Rows | #Files | Size | Format | ++-------+-------+--------+------+---------+ +| 2000 | -1 | 0 | 0B | TEXT | +| 2004 | -1 | 0 | 0B | TEXT | +| 2008 | -1 | 0 | 0B | TEXT | +| 2010 | -1 | 0 | 0B | TEXT | +| 2011 | 4 | 1 | 22B | TEXT | +| 2012 | 4 | 1 | 22B | TEXT | +| 2013 | 1 | 1 | 231B | PARQUET | +| Total | 9 | 3 | 275B | | ++-------+-------+--------+------+---------+ +Returned 8 row(s) in 0.02s +[localhost:21000] > show column stats census; ++--------+----------+------------------+--------+----------+----------+ +| Column | Type | #Distinct Values | #Nulls | Max Size | Avg Size | ++--------+----------+------------------+--------+----------+----------+ +| name | STRING | 4 | -1 | 5 | 4.5 | +| year | SMALLINT | 7 | -1 | 2 | 2 | ++--------+----------+------------------+--------+----------+----------+ +Returned 2 row(s) in 0.02s</codeblock> + + <p rev="1.4.0"> + For examples showing how some queries work differently when statistics are available, see + <xref href="impala_perf_joins.xml#perf_joins_examples"/>. You can see how Impala executes a query + differently in each case by observing the <codeph>EXPLAIN</codeph> output before and after collecting + statistics. Measure the before and after query times, and examine the throughput numbers in before and + after <codeph>SUMMARY</codeph> or <codeph>PROFILE</codeph> output, to verify how much the improved plan + speeds up performance. + </p> + </conbody> + </concept> +</concept>
http://git-wip-us.apache.org/repos/asf/incubator-impala/blob/3be0f122/docs/topics/impala_perf_testing.xml ---------------------------------------------------------------------- diff --git a/docs/topics/impala_perf_testing.xml b/docs/topics/impala_perf_testing.xml new file mode 100644 index 0000000..d621556 --- /dev/null +++ b/docs/topics/impala_perf_testing.xml @@ -0,0 +1,175 @@ +<?xml version="1.0" encoding="UTF-8"?> +<!DOCTYPE concept PUBLIC "-//OASIS//DTD DITA Concept//EN" "concept.dtd"> +<concept id="performance_testing"> + + <title>Testing Impala Performance</title> + <prolog> + <metadata> + <data name="Category" value="Impala"/> + <data name="Category" value="Performance"/> + <data name="Category" value="Troubleshooting"/> + <data name="Category" value="Proof of Concept"/> + <data name="Category" value="Logs"/> + <data name="Category" value="Administrators"/> + <data name="Category" value="Developers"/> + <data name="Category" value="Data Analysts"/> + <!-- Should reorg this topic to use nested topics, not sections. Some keywords like 'logs' buried in section titles. --> + <data name="Category" value="Sectionated Pages"/> + </metadata> + </prolog> + + <conbody> + + <p> + Test to ensure that Impala is configured for optimal performance. If you have installed Impala without + Cloudera Manager, complete the processes described in this topic to help ensure a proper configuration. Even + if you installed Impala with Cloudera Manager, which automatically applies appropriate configurations, these + procedures can be used to verify that Impala is set up correctly. + </p> + + <section id="checking_config_performance"> + + <title>Checking Impala Configuration Values</title> + + <p> + You can inspect Impala configuration values by connecting to your Impala server using a browser. + </p> + + <p> + <b>To check Impala configuration values:</b> + </p> + + <ol> + <li> + Use a browser to connect to one of the hosts running <codeph>impalad</codeph> in your environment. + Connect using an address of the form + <codeph>http://<varname>hostname</varname>:<varname>port</varname>/varz</codeph>. + <note> + In the preceding example, replace <codeph>hostname</codeph> and <codeph>port</codeph> with the name and + port of your Impala server. The default port is 25000. + </note> + </li> + + <li> + Review the configured values. + <p> + For example, to check that your system is configured to use block locality tracking information, you + would check that the value for <codeph>dfs.datanode.hdfs-blocks-metadata.enabled</codeph> is + <codeph>true</codeph>. + </p> + </li> + </ol> + + <p id="p_31"> + <b>To check data locality:</b> + </p> + + <ol> + <li> + Execute a query on a dataset that is available across multiple nodes. For example, for a table named + <codeph>MyTable</codeph> that has a reasonable chance of being spread across multiple DataNodes: +<codeblock>[impalad-host:21000] > SELECT COUNT (*) FROM MyTable</codeblock> + </li> + + <li> + After the query completes, review the contents of the Impala logs. You should find a recent message + similar to the following: +<codeblock>Total remote scan volume = 0</codeblock> + </li> + </ol> + + <p> + The presence of remote scans may indicate <codeph>impalad</codeph> is not running on the correct nodes. + This can be because some DataNodes do not have <codeph>impalad</codeph> running or it can be because the + <codeph>impalad</codeph> instance that is starting the query is unable to contact one or more of the + <codeph>impalad</codeph> instances. + </p> + + <p> + <b>To understand the causes of this issue:</b> + </p> + + <ol> + <li> + Connect to the debugging web server. By default, this server runs on port 25000. This page lists all + <codeph>impalad</codeph> instances running in your cluster. If there are fewer instances than you expect, + this often indicates some DataNodes are not running <codeph>impalad</codeph>. Ensure + <codeph>impalad</codeph> is started on all DataNodes. + </li> + + <li> + <!-- To do: + There are other references to this tip about the "Impala daemon's hostname" elsewhere. Could reconcile, conref, or link. + --> + If you are using multi-homed hosts, ensure that the Impala daemon's hostname resolves to the interface on + which <codeph>impalad</codeph> is running. The hostname Impala is using is displayed when + <codeph>impalad</codeph> starts. To explicitly set the hostname, use the <codeph>--hostname</codeph>Â flag. + </li> + + <li> + Check that <codeph>statestored</codeph> is running as expected. Review the contents of the state store + log to ensure all instances of <codeph>impalad</codeph> are listed as having connected to the state + store. + </li> + </ol> + </section> + + <section id="checking_config_logs"> + + <title>Reviewing Impala Logs</title> + + <p> + You can review the contents of the Impala logs for signs that short-circuit reads or block location + tracking are not functioning. Before checking logs, execute a simple query against a small HDFS dataset. + Completing a query task generates log messages using current settings. Information on starting Impala and + executing queries can be found in <xref href="impala_processes.xml#processes"/> and + <xref href="impala_impala_shell.xml#impala_shell"/>. Information on logging can be found in + <xref href="impala_logging.xml#logging"/>. Log messages and their interpretations are as follows: + </p> + + <table> + <tgroup cols="2"> + <colspec colname="1" colwidth="30*"/> + <colspec colname="2" colwidth="10*"/> + <thead> + <row> + <entry> + Log Message + </entry> + <entry> + Interpretation + </entry> + </row> + </thead> + <tbody> + <row> + <entry> + <p> +<pre>Unknown disk id. This will negatively affect performance. Check your hdfs settings to enable block location metadata +</pre> + </p> + </entry> + <entry> + <p> + Tracking block locality is not enabled. + </p> + </entry> + </row> + <row> + <entry> + <p> +<pre>Unable to load native-hadoop library for your platform... using builtin-java classes where applicable</pre> + </p> + </entry> + <entry> + <p> + Native checksumming is not enabled. + </p> + </entry> + </row> + </tbody> + </tgroup> + </table> + </section> + </conbody> +</concept> http://git-wip-us.apache.org/repos/asf/incubator-impala/blob/3be0f122/docs/topics/impala_performance.xml ---------------------------------------------------------------------- diff --git a/docs/topics/impala_performance.xml b/docs/topics/impala_performance.xml new file mode 100644 index 0000000..e58270e --- /dev/null +++ b/docs/topics/impala_performance.xml @@ -0,0 +1,191 @@ +<?xml version="1.0" encoding="UTF-8"?> +<!DOCTYPE concept PUBLIC "-//OASIS//DTD DITA Concept//EN" "concept.dtd"> +<concept id="performance"> + + <title>Tuning Impala for Performance</title> + <titlealts audience="PDF"><navtitle>Performance Tuning</navtitle></titlealts> + <prolog> + <metadata> + <data name="Category" value="Impala"/> + <data name="Category" value="Performance"/> + <data name="Category" value="Databases"/> + <data name="Category" value="SQL"/> + <data name="Category" value="Querying"/> + <data name="Category" value="Developers"/> + <!-- Like Impala Administration, this page has a fair bit of info already, but it could benefit from wiki-style embedded of intro text from those other pages. --> + <data name="Category" value="Stub Pages"/> + </metadata> + </prolog> + + <conbody> + + <p> + The following sections explain the factors affecting the performance of Impala features, and procedures for + tuning, monitoring, and benchmarking Impala queries and other SQL operations. + </p> + + <p> + This section also describes techniques for maximizing Impala scalability. Scalability is tied to performance: + it means that performance remains high as the system workload increases. For example, reducing the disk I/O + performed by a query can speed up an individual query, and at the same time improve scalability by making it + practical to run more queries simultaneously. Sometimes, an optimization technique improves scalability more + than performance. For example, reducing memory usage for a query might not change the query performance much, + but might improve scalability by allowing more Impala queries or other kinds of jobs to run at the same time + without running out of memory. + </p> + + <note> + <p> + Before starting any performance tuning or benchmarking, make sure your system is configured with all the + recommended minimum hardware requirements from <xref href="impala_prereqs.xml#prereqs_hardware"/> and + software settings from <xref href="impala_config_performance.xml#config_performance"/>. + </p> + </note> + + <ul> + <li> + <xref href="impala_partitioning.xml#partitioning"/>. This technique physically divides the data based on + the different values in frequently queried columns, allowing queries to skip reading a large percentage of + the data in a table. + </li> + + <li> + <xref href="impala_perf_joins.xml#perf_joins"/>. Joins are the main class of queries that you can tune at + the SQL level, as opposed to changing physical factors such as the file format or the hardware + configuration. The related topics <xref href="impala_perf_stats.xml#perf_column_stats"/> and + <xref href="impala_perf_stats.xml#perf_table_stats"/> are also important primarily for join performance. + </li> + + <li> + <xref href="impala_perf_stats.xml#perf_table_stats"/> and + <xref href="impala_perf_stats.xml#perf_column_stats"/>. Gathering table and column statistics, using the + <codeph>COMPUTE STATS</codeph> statement, helps Impala automatically optimize the performance for join + queries, without requiring changes to SQL query statements. (This process is greatly simplified in Impala + 1.2.2 and higher, because the <codeph>COMPUTE STATS</codeph> statement gathers both kinds of statistics in + one operation, and does not require any setup and configuration as was previously necessary for the + <codeph>ANALYZE TABLE</codeph> statement in Hive.) + </li> + + <li> + <xref href="impala_perf_testing.xml#performance_testing"/>. Do some post-setup testing to ensure Impala is + using optimal settings for performance, before conducting any benchmark tests. + </li> + + <li> + <xref href="impala_perf_benchmarking.xml#perf_benchmarks"/>. The configuration and sample data that you use + for initial experiments with Impala is often not appropriate for doing performance tests. + </li> + + <li> + <xref href="impala_perf_resources.xml#mem_limits"/>. The more memory Impala can utilize, the better query + performance you can expect. In a cluster running other kinds of workloads as well, you must make tradeoffs + to make sure all Hadoop components have enough memory to perform well, so you might cap the memory that + Impala can use. + </li> + + <li rev="1.2" audience="Cloudera"> + <xref href="impala_perf_hdfs_caching.xml#hdfs_caching"/>. Impala can use the HDFS caching feature to pin + frequently accessed data in memory, reducing disk I/O. + </li> + + <li rev="2.2.0"> + <xref href="impala_s3.xml#s3"/>. Queries against data stored in the Amazon Simple Storage Service (S3) + have different performance characteristics than when the data is stored in HDFS. + </li> + </ul> + + <p outputclass="toc"/> + + <p conref="../shared/impala_common.xml#common/cookbook_blurb"/> + + </conbody> + +<!-- Empty/hidden stub sections that might be worth expanding later. --> + + <concept id="perf_network" audience="Cloudera"> + + <title>Network Traffic</title> + + <conbody/> + </concept> + + <concept id="perf_partition_schema" audience="Cloudera"> + + <title>Designing Partitioned Tables</title> + + <conbody/> + </concept> + + <concept id="perf_partition_query" audience="Cloudera"> + + <title>Queries on Partitioned Tables</title> + + <conbody/> + </concept> + + <concept id="perf_monitoring" audience="Cloudera"> + + <title>Monitoring Performance through the Impala Web Interface</title> + <prolog> + <metadata> + <data name="Category" value="Monitoring"/> + </metadata> + </prolog> + + <conbody/> + </concept> + + <concept id="perf_query_coord" audience="Cloudera"> + + <title>Query Coordination</title> + + <conbody/> + </concept> + + <concept id="perf_bottlenecks" audience="Cloudera"> + + <title>Performance Bottlenecks</title> + + <conbody/> + </concept> + + <concept id="perf_long_queries" audience="Cloudera"> + + <title>Managing Long-Running Queries</title> + + <conbody/> + </concept> + + <concept id="perf_load" audience="Cloudera"> + + <title>Performance Considerations for Loading Data</title> + + <conbody/> + </concept> + + <concept id="perf_file_formats" audience="Cloudera"> + + <title>Performance Considerations for File Formats</title> + + <conbody/> + </concept> + + <concept id="perf_compression" audience="Cloudera"> + + <title>Performance Considerations for Compression</title> + <prolog> + <metadata> + <data name="Category" value="Compression"/> + </metadata> + </prolog> + + <conbody/> + </concept> + + <concept id="perf_codegen" audience="Cloudera"> + + <title>Native Code Generation</title> + + <conbody/> + </concept> +</concept> http://git-wip-us.apache.org/repos/asf/incubator-impala/blob/3be0f122/docs/topics/impala_planning.xml ---------------------------------------------------------------------- diff --git a/docs/topics/impala_planning.xml b/docs/topics/impala_planning.xml new file mode 100644 index 0000000..f103ab8 --- /dev/null +++ b/docs/topics/impala_planning.xml @@ -0,0 +1,30 @@ +<?xml version="1.0" encoding="UTF-8"?> +<!DOCTYPE concept PUBLIC "-//OASIS//DTD DITA Concept//EN" "concept.dtd"> +<concept id="planning"> + + <title>Planning for Impala Deployment</title> + <titlealts audience="PDF"><navtitle>Deployment Planning</navtitle></titlealts> + <prolog> + <metadata> + <data name="Category" value="Impala"/> + <data name="Category" value="Deploying"/> + <data name="Category" value="Planning"/> + <data name="Category" value="Proof of Concept"/> + <data name="Category" value="Administrators"/> + <data name="Category" value="Developers"/> + <data name="Category" value="Stub Pages"/> + </metadata> + </prolog> + + <conbody> + + <p> + <indexterm audience="Cloudera">planning</indexterm> + Before you set up Impala in production, do some planning to make sure that your hardware setup has sufficient + capacity, that your cluster topology is optimal for Impala queries, and that your schema design and ETL + processes follow the best practices for Impala. + </p> + + <p outputclass="toc"/> + </conbody> +</concept>
