Introduction to Statistics

TiDB uses statistics to decide which index to choose. The tidb_analyze_version variable controls the statistics collected by TiDB. Currently, two versions of statistics are supported: tidb_analyze_version = 1 (by default) and tidb_analyze_version = 2. These two versions include different information in TiDB:

Compared to Version 1, Version 2 statistics avoids the potential inaccuracy caused by hash collision when the data volume is huge. It also increases the estimate precision in most scenarios.

This document briefly introduces the histogram, Count-Min Sketch, and Top-N, and details the collection and maintenance of statistics.

Histogram

A histogram is an approximate representation of the distribution of data. It divides the entire range of values into a series of buckets, and uses simple data to describe each bucket, such as the number of values ​​falling in the bucket. In TiDB, an equal-depth histogram is created for the specific columns of each table. The equal-depth histogram can be used to estimate the interval query.

Here "equal-depth" means that the number of values ​​falling into each bucket is as equal as possible. For example, for a given set {1.6, 1.9, 1.9, 2.0, 2.4, 2.6, 2.7, 2.7, 2.8, 2.9, 3.4, 3.5}, you want to generate 4 buckets. The equal-depth histogram is as follows. It contains four buckets [1.6, 1.9], [2.0, 2.6], [2.7, 2.8], [2.9, 3.5]. The bucket depth is 3.

For details about the parameter that determines the upper limit to the number of histogram buckets, refer to Manual Collection. When the number of buckets is larger, the accuracy of the histogram is higher; however, higher accuracy is at the cost of the usage of memory resources. You can adjust this number appropriately according to the actual scenario.

Count-Min Sketch

Count-Min Sketch is a hash structure. When an equivalence query contains a = 1 or IN query (for example, a in (1, 2, 3)), TiDB uses this data structure for estimation.

A hash collision might occur since Count-Min Sketch is a hash structure. In the EXPLAIN statement, if the estimate of the equivalent query deviates greatly from the actual value, it can be considered that a larger value and a smaller value have been hashed together. In this case, you can take one of the following ways to avoid the hash collision:

• Modify the WITH NUM TOPN parameter. TiDB stores the high-frequency (top x) data separately, with the other data stored in Count-Min Sketch. Therefore, to prevent a larger value and a smaller value from being hashed together, you can increase the value of WITH NUM TOPN. In TiDB, its default value is 20. The maximum value is 1024. For more information about this parameter, see Full Collection.
• Modify two parameters WITH NUM CMSKETCH DEPTH and WITH NUM CMSKETCH WIDTH. Both affect the number of hash buckets and the collision probability. You can increase the values of the two parameters appropriately according to the actual scenario to reduce the probability of hash collision, but at the cost of higher memory usage of statistics. In TiDB, the default value of WITH NUM CMSKETCH DEPTH is 5, and the default value of WITH NUM CMSKETCH WIDTH is 2048. For more information about the two parameters, see Full Collection.

Top-N values

Top-N values are values with the top N occurrences in a column or index. TiDB records the values and occurences of Top-N values.

Collect statistics

Manual collection

You can run the ANALYZE statement to collect statistics.

Full collection

Note:

The execution time of ANALYZE TABLE in TiDB is longer than that in MySQL or InnoDB. In InnoDB, only a small number of pages are sampled, while in TiDB a comprehensive set of statistics is completely rebuilt. Scripts that were written for MySQL may naively expect ANALYZE TABLE will be a short-lived operation.

For quicker analysis, you can set tidb_enable_fast_analyze to 1 to enable the Quick Analysis feature. The default value for this parameter is 0.

After Quick Analysis is enabled, TiDB randomly samples approximately 10,000 rows of data to build statistics. Therefore, in the case of uneven data distribution or a relatively small amount of data, the accuracy of statistical information is relatively poor. It might lead to poor execution plans, such as choosing the wrong index. If the execution time of the normal ANALYZE statement is acceptable, it is recommended to disable the Quick Analysis feature.

You can perform full collection using the following syntax.

• To collect statistics of all the tables in TableNameList:

  Copy

  ANALYZE TABLE TableNameList [WITH NUM BUCKETS|TOPN|CMSKETCH DEPTH|CMSKETCH WIDTH|SAMPLES];

• WITH NUM BUCKETS specifies the maximum number of buckets in the generated histogram.

• WITH NUM TOPN specifies the maximum number of the generated TOPNs.

• WITH NUM CMSKETCH DEPTH specifies the depth of the CM Sketch.

• WITH NUM CMSKETCH WIDTH specifies the width of the CM Sketch.

• WITH NUM SAMPLES specifies the number of samples.

• To collect statistics of the index columns on all IndexNameLists in TableName:

  Copy

  ANALYZE TABLE TableName INDEX [IndexNameList] [WITH NUM BUCKETS|TOPN|CMSKETCH DEPTH|CMSKETCH WIDTH|SAMPLES];

  The statement collects statistics of all index columns when IndexNameList is empty.

• To collect statistics of partition in all PartitionNameLists in TableName:

  Copy

  ANALYZE TABLE TableName PARTITION PartitionNameList [WITH NUM BUCKETS|TOPN|CMSKETCH DEPTH|CMSKETCH WIDTH|SAMPLES];

• To collect statistics of index columns for the partitions in all PartitionNameLists in TableName:

  Copy

  ANALYZE TABLE TableName PARTITION PartitionNameList [IndexNameList] [WITH NUM BUCKETS|TOPN|CMSKETCH DEPTH|CMSKETCH WIDTH|SAMPLES];

Incremental collection

To improve the speed of analysis after full collection, incremental collection could be used to analyze the newly added sections in monotonically non-decreasing columns such as time columns.

Note:

• Currently, the incremental collection is only provided for index.
• When using the incremental collection, you must ensure that only INSERT operations exist on the table, and that the newly inserted value on the index column is monotonically non-decreasing. Otherwise, the statistical information might be inaccurate, affecting the TiDB optimizer to select an appropriate execution plan.

You can perform incremental collection using the following syntax.

• To incrementally collect statistics for index columns in all IndexNameLists in TableName:

  Copy

  ANALYZE INCREMENTAL TABLE TableName INDEX [IndexNameList] [WITH NUM BUCKETS|TOPN|CMSKETCH DEPTH|CMSKETCH WIDTH|SAMPLES];

• To incrementally collect statistics of index columns for partitions in all PartitionNameLists in TableName:

  Copy

  ANALYZE INCREMENTAL TABLE TableName PARTITION PartitionNameList INDEX [IndexNameList] [WITH NUM BUCKETS|TOPN|CMSKETCH DEPTH|CMSKETCH WIDTH|SAMPLES];

Automatic update

For the INSERT, DELETE, or UPDATE statements, TiDB automatically updates the number of rows and updated rows. TiDB persists this information regularly and the update cycle is 20 * stats-lease. The default value of stats-lease is 3s. If you specify the value as 0, it does not update automatically.

Three system variables related to automatic update of statistics are as follows:

When the ratio of the number of modified rows to the total number of rows of tbl in a table is greater than tidb_auto_analyze_ratio, and the current time is between tidb_auto_analyze_start_time and tidb_auto_analyze_end_time, TiDB executes the ANALYZE TABLE tbl statement in the background to automatically update the statistics of this table.

When the query is executed, TiDB collects feedback with the probability of feedback-probability and uses it to update the histogram and Count-Min Sketch. You can modify the value of feedback-probability in the configuration file. The default value is 0.05. You can set the value to 0.0 to disable this feature.

Note:

If you set the value of feedback-probability to 0 in the configuration file, a failure will occur and an error will be reported. To disable feedback-probability, you need to set the value to 0.0.

Control ANALYZE concurrency

When you run the ANALYZE statement, you can adjust the concurrency using the following parameters, to control its effect on the system.

tidb_build_stats_concurrency

Currently, when you run the ANALYZE statement, the task is divided into multiple small tasks. Each task only works on one column or index. You can use the tidb_build_stats_concurrency parameter to control the number of simultaneous tasks. The default value is 4.

tidb_distsql_scan_concurrency

When you analyze regular columns, you can use the tidb_distsql_scan_concurrency parameter to control the number of Region to be read at one time. The default value is 15.

tidb_index_serial_scan_concurrency

When you analyze index columns, you can use the tidb_index_serial_scan_concurrency parameter to control the number of Region to be read at one time. The default value is 1.

View ANALYZE state

When executing the ANALYZE statement, you can view the current state of ANALYZE using the following SQL statement:

SHOW ANALYZE STATUS [ShowLikeOrWhere]

This statement returns the state of ANALYZE. You can use ShowLikeOrWhere to filter the information you need.

Currently, the SHOW ANALYZE STATUS statement returns the following 7 columns:

View statistics

You can view the statistics status using the following statements.

Metadata of tables

You can use the SHOW STATS_META statement to view the total number of rows and the number of updated rows.

The syntax of ShowLikeOrWhereOpt is as follows:

SHOW STATS_META [ShowLikeOrWhere]

Currently, the SHOW STATS_META statement returns the following 6 columns:

Note:

When TiDB automatically updates the total number of rows and the number of modified rows according to DML statements, update_time is also updated. Therefore, update_time does not necessarily indicate the last time when the ANALYZE statement is executed.

Health state of tables

You can use the SHOW STATS_HEALTHY statement to check the health state of tables and roughly estimate the accuracy of the statistics. When modify_count >= row_count, the health state is 0; when modify_count < row_count, the health state is (1 - modify_count/row_count) * 100.

The synopsis of SHOW STATS_HEALTHY is:

and the synopsis of the ShowLikeOrWhereOpt part is:

Currently, the SHOW STATS_HEALTHY statement returns the following 4 columns:

Metadata of columns

You can use the SHOW STATS_HISTOGRAMS statement to view the number of different values and the number of NULL in all the columns.

Syntax as follows:

SHOW STATS_HISTOGRAMS [ShowLikeOrWhere]

This statement returns the number of different values and the number of NULL in all the columns. You can use ShowLikeOrWhere to filter the information you need.

Currently, the SHOW STATS_HISTOGRAMS statement returns the following 10 columns:

Buckets of histogram

You can use the SHOW STATS_BUCKETS statement to view each bucket of the histogram.

The syntax is as follows:

SHOW STATS_BUCKETS [ShowLikeOrWhere]

The diagram is as follows:

This statement returns information about all the buckets. You can use ShowLikeOrWhere to filter the information you need.

Currently, the SHOW STATS_BUCKETS statement returns the following 11 columns:

Top-N information

You can use the SHOW STATS_TOPN statement to view the Top-N information currently collected by TiDB.

The syntax is as follows:

SHOW STATS_TOPN [ShowLikeOrWhere];

Currently, the SHOW STATS_TOPN statement returns the following 7 columns:

Delete statistics

You can run the DROP STATS statement to delete statistics.

Syntax as follows:

DROP STATS TableName

The statement deletes statistics of all the tables in TableName.

Import and export statistics

Export statistics

The interface to export statistics is as follows:

• To obtain the JSON format statistics of the ${table_name} table in the ${db_name} database:

  Copy

  http://${tidb-server-ip}:${tidb-server-status-port}/stats/dump/${db_name}/${table_name}

• To obtain the JSON format statistics of the ${table_name} table in the ${db_name} database at specific time:

  Copy

  http://${tidb-server-ip}:${tidb-server-status-port}/stats/dump/${db_name}/${table_name}/${yyyyMMddHHmmss}

Import statistics

Generally, the imported statistics refer to the JSON file obtained using the export interface.

Syntax:

LOAD STATS 'file_name'

file_name is the file name of the statistics to be imported.

See also

• DROP STATS