diff --git a/TOC-tidb-cloud.md b/TOC-tidb-cloud.md index 55eab66ef9789..37c66a6a0125b 100644 --- a/TOC-tidb-cloud.md +++ b/TOC-tidb-cloud.md @@ -839,6 +839,7 @@ - [Table Filter](/table-filter.md) - [URI Formats of External Storage Services](/external-storage-uri.md) - [DDL Execution Principles and Best Practices](/ddl-introduction.md) + - [`ANALYZE` Embedded in DDL Statements](/ddl_embedded_analyze.md) - [Batch Processing](/batch-processing.md) - [Troubleshoot Inconsistency Between Data and Indexes](/troubleshoot-data-inconsistency-errors.md) - [Notifications](/tidb-cloud/notifications.md) diff --git a/TOC.md b/TOC.md index d7b7464f48993..357996a99e564 100644 --- a/TOC.md +++ b/TOC.md @@ -1088,7 +1088,8 @@ - [URI Formats of External Storage Services](/external-storage-uri.md) - [TiDB Workload Repository](/workload-repository.md) - [Interaction Test on Online Workloads and `ADD INDEX` Operations](/benchmark/online-workloads-and-add-index-operations.md) -- FAQs + - [`ANALYZE` Embedded in DDL Statements](/ddl_embedded_analyze.md) + FAQs - [FAQ Summary](/faq/faq-overview.md) - [TiDB FAQs](/faq/tidb-faq.md) - [SQL FAQs](/faq/sql-faq.md) diff --git a/ddl_embedded_analyze.md b/ddl_embedded_analyze.md new file mode 100644 index 0000000000000..9de981b1c2467 --- /dev/null +++ b/ddl_embedded_analyze.md @@ -0,0 +1,177 @@ +--- +title: `ANALYZE` Embedded in DDL Statements +summary: This document describes the `ANALYZE` feature embedded in DDL statements for newly created or reorganized indexes, which ensures that statistics for new indexes are updated promptly. +--- + +# `ANALYZE` Embedded in DDL StatementsIntroduced in v8.5.4 and v9.0.0 + +This document describes the `ANALYZE` feature embedded in the following two types of DDL statements: + +- DDL statements that create new indexes: [`ADD INDEX`](/sql-statements/sql-statement-add-index.md) +- DDL statements that reorganize existing indexes: [`MODIFY COLUMN`](/sql-statements/sql-statement-modify-column.md) and [`CHANGE COLUMN`](/sql-statements/sql-statement-change-column.md) + +When this feature is enabled, TiDB automatically runs an `ANALYZE` (statistics collection) operation before the new or reorganized index becomes visible to users. This prevents inaccurate optimizer estimates and potential plan changes caused by temporarily unavailable statistics after index creation or reorganization. + +## Usage scenarios + +In scenarios where DDL operations alternately add or modify indexes, existing stable queries might suffer from estimation bias because the new index lacks statistics, causing the optimizer to choose suboptimal plans. For more information, see [Issue #57948](https://github.com/pingcap/tidb/issues/57948). + +For example: + +```sql +CREATE TABLE t (a INT, b INT); +INSERT INTO t VALUES (1, 1), (2, 2), (3, 3); +INSERT INTO t SELECT * FROM t; -- * N times + +ALTER TABLE t ADD INDEX idx_a (a); + +EXPLAIN SELECT * FROM t WHERE a > 4; +``` + +``` ++-------------------------+-----------+-----------+---------------+--------------------------------+ +| id | estRows | task | access object | operator info | ++-------------------------+-----------+-----------+---------------+--------------------------------+ +| TableReader_8 | 131072.00 | root | | data:Selection_7 | +| └─Selection_7 | 131072.00 | cop[tikv] | | gt(test.t.a, 4) | +| └─TableFullScan_6 | 393216.00 | cop[tikv] | table:t | keep order:false, stats:pseudo | ++-------------------------+-----------+-----------+---------------+--------------------------------+ +3 rows in set (0.002 sec) +``` + +In the preceding plan, because the newly created index has no statistics yet, TiDB can only rely on heuristic rules for path estimation. Unless the index access path requires no table lookup and has a significantly lower cost, the optimizer tends to choose the more stable existing path. In the preceding example, it chooses a full table scan. However, from the data distribution perspective, `t.a > 4` actually returns 0 rows. If the new index `idx_a` were used, the query could quickly locate relevant rows and avoid the full table scan. In this example, because statistics are not promptly collected after the DDL creates the index, the generated plan is not optimal, but the optimizer continues to use the original plan so query performance does not sharply regress. However, according to [Issue #57948](https://github.com/pingcap/tidb/issues/57948), in some cases heuristics might cause an unreasonable comparison between old and new indexes, pruning the index that the original plan relies on and ultimately falling back to a full table scan. + +Starting from v8.5.0, TiDB has improved heuristic comparisons between indexes and behaviors when statistics are missing. Still, in some complex scenarios, embedding `ANALYZE` in DDL is the best way to prevent plan changes. You can control whether to run embedded `ANALYZE` during index creation or reorganization with the system variable [`tidb_stats_update_during_ddl`](/system-variables.md#tidb_stats_update_during_ddl-new-in-v854-and-v900). The default value is `OFF`. + +## `ADD INDEX` DDL + +When `tidb_stats_update_during_ddl` is `ON`, executing [`ADD INDEX`](/sql-statements/sql-statement-add-index.md) automatically runs an embedded `ANALYZE` operation after the Reorg phase finishes. This `ANALYZE` operation collects statistics for the newly created index before the index becomes visible to users, and then `ADD INDEX` proceeds with its remaining phases. + +Considering that `ANALYZE` can take time, TiDB sets a timeout threshold based on the execution time of the first Reorg. If `ANALYZE` times out, `ADD INDEX` stops waiting synchronously for `ANALYZE` to finish and continues the subsequent process, making the index visible earlier to users. This means the index statistics will be updated after `ANALYZE` completes asynchronously. + +For example: + +```sql +CREATE TABLE t (a INT, b INT, c INT); +Query OK, 0 rows affected (0.011 sec) + +INSERT INTO t VALUES (1, 1, 1), (2, 2, 2), (3, 3, 3); +Query OK, 3 rows affected (0.003 sec) +Records: 3 Duplicates: 0 Warnings: 0 + +SET @@tidb_stats_update_during_ddl = 1; +Query OK, 0 rows affected (0.001 sec) + +ALTER TABLE t ADD INDEX idx (a, b); +Query OK, 0 rows affected (0.049 sec) +``` + +```sql +EXPLAIN SELECT a FROM t WHERE a > 1; +``` + +``` ++------------------------+---------+-----------+--------------------------+----------------------------------+ +| id | estRows | task | access object | operator info | ++------------------------+---------+-----------+--------------------------+----------------------------------+ +| IndexReader_7 | 4.00 | root | | index:IndexRangeScan_6 | +| └─IndexRangeScan_6 | 4.00 | cop[tikv] | table:t, index:idx(a, b) | range:(1,+inf], keep order:false | ++------------------------+---------+-----------+--------------------------+----------------------------------+ +2 rows in set (0.002 sec) +``` + +```sql +SHOW STATS_HISTOGRAMS WHERE table_name = "t"; +``` + +``` ++---------+------------+----------------+-------------+----------+---------------------+----------------+------------+--------------+-------------+-------------+-----------------+----------------+----------------+---------------+ +| Db_name | Table_name | Partition_name | Column_name | Is_index | Update_time | Distinct_count | Null_count | Avg_col_size | Correlation | Load_status | Total_mem_usage | Hist_mem_usage | Topn_mem_usage | Cms_mem_usage | ++---------+------------+----------------+-------------+----------+---------------------+----------------+------------+--------------+-------------+-------------+-----------------+----------------+----------------+---------------+ +| test | t | | a | 0 | 2025-10-30 20:17:57 | 3 | 0 | 0.5 | 1 | allLoaded | 155 | 0 | 155 | 0 | +| test | t | | idx | 1 | 2025-10-30 20:17:57 | 3 | 0 | 0 | 0 | allLoaded | 182 | 0 | 182 | 0 | ++---------+------------+----------------+-------------+----------+---------------------+----------------+------------+--------------+-------------+-------------+-----------------+----------------+----------------+---------------+ +2 rows in set (0.013 sec) +``` + +```sql +ADMIN SHOW DDL JOBS 1; +``` + +``` ++--------+---------+--------------------------+---------------+----------------------+-----------+----------+-----------+----------------------------+----------------------------+----------------------------+---------+----------------------------------------+ +| JOB_ID | DB_NAME | TABLE_NAME | JOB_TYPE | SCHEMA_STATE | SCHEMA_ID | TABLE_ID | ROW_COUNT | CREATE_TIME | START_TIME | END_TIME | STATE | COMMENTS | ++--------+---------+--------------------------+---------------+----------------------+-----------+----------+-----------+----------------------------+----------------------------+----------------------------+---------+----------------------------------------+ +| 151 | test | t | add index | write reorganization | 2 | 148 | 6291456 | 2025-10-29 00:14:47.181000 | 2025-10-29 00:14:47.183000 | NULL | running | analyzing, txn-merge, max_node_count=3 | ++--------+---------+--------------------------+---------------+----------------------+-----------+----------+-----------+----------------------------+----------------------------+----------------------------+---------+----------------------------------------+ +1 rows in set (0.001 sec) +``` + +From the `ADD INDEX` example, when `tidb_stats_update_during_ddl` is `ON`, you can see that after the execution of the `ADD INDEX` DDL statement, the subsequent `EXPLAIN` output shows that statistics for the index `idx` have been automatically collected and loaded into memory (you can verify it by executing `SHOW STATS_HISTOGRAMS`). As a result, the optimizer can immediately use these statistics for range scans. If index creation or reorganization and `ANALYZE` take a long time, you can check the DDL job status by executing `ADMIN SHOW DDL JOBS`. When the `COMMENTS` column in the output contains `analyzing`, it means that the DDL job is collecting statistics. + +## DDL for reorganizing existing indexes + +When `tidb_stats_update_during_ddl` is `ON`, executing [`MODIFY COLUMN`](/sql-statements/sql-statement-modify-column.md) or [`CHANGE COLUMN`](/sql-statements/sql-statement-change-column.md) that reorganizes an index will also run an embedded `ANALYZE` operation after the Reorg phase completes. The mechanism is the same as for `ADD INDEX`: + +- Start collecting statistics before the index becomes visible. +- If `ANALYZE` times out, [`MODIFY COLUMN`](/sql-statements/sql-statement-modify-column.md) and [`CHANGE COLUMN`](/sql-statements/sql-statement-change-column.md) stops waiting synchronously for `ANALYZE` to finish and continues the subsequent process, making the index visible earlier to users. This means that the index statistics will be updated when `ANALYZE` finishes asynchronously. + +For example: + +```sql +CREATE TABLE s (a VARCHAR(10), INDEX idx (a)); +Query OK, 0 rows affected (0.012 sec) + +INSERT INTO s VALUES (1), (2), (3); +Query OK, 3 rows affected (0.003 sec) +Records: 3 Duplicates: 0 Warnings: 0 + +SET @@tidb_stats_update_during_ddl = 1; +Query OK, 0 rows affected (0.001 sec) + +ALTER TABLE s MODIFY COLUMN a INT; +Query OK, 0 rows affected (0.056 sec) + +EXPLAIN SELECT * FROM s WHERE a > 1; +``` + +``` ++------------------------+---------+-----------+-----------------------+----------------------------------+ +| id | estRows | task | access object | operator info | ++------------------------+---------+-----------+-----------------------+----------------------------------+ +| IndexReader_7 | 2.00 | root | | index:IndexRangeScan_6 | +| └─IndexRangeScan_6 | 2.00 | cop[tikv] | table:s, index:idx(a) | range:(1,+inf], keep order:false | ++------------------------+---------+-----------+-----------------------+----------------------------------+ +2 rows in set (0.005 sec) +``` + +```sql +SHOW STATS_HISTOGRAMS WHERE table_name = "s"; +``` + +``` ++---------+------------+----------------+-------------+----------+---------------------+----------------+------------+--------------+-------------+-------------+-----------------+----------------+----------------+---------------+ +| Db_name | Table_name | Partition_name | Column_name | Is_index | Update_time | Distinct_count | Null_count | Avg_col_size | Correlation | Load_status | Total_mem_usage | Hist_mem_usage | Topn_mem_usage | Cms_mem_usage | ++---------+------------+----------------+-------------+----------+---------------------+----------------+------------+--------------+-------------+-------------+-----------------+----------------+----------------+---------------+ +| test | s | | a | 0 | 2025-10-30 20:10:18 | 3 | 0 | 2 | 1 | allLoaded | 158 | 0 | 158 | 0 | +| test | s | | a | 0 | 2025-10-30 20:10:18 | 3 | 0 | 1 | 1 | allLoaded | 155 | 0 | 155 | 0 | +| test | s | | idx | 1 | 2025-10-30 20:10:18 | 3 | 0 | 0 | 0 | allLoaded | 158 | 0 | 158 | 0 | +| test | s | | idx | 1 | 2025-10-30 20:10:18 | 3 | 0 | 0 | 0 | allLoaded | 155 | 0 | 155 | 0 | ++---------+------------+----------------+-------------+----------+---------------------+----------------+------------+--------------+-------------+-------------+-----------------+----------------+----------------+---------------+ +4 rows in set (0.008 sec) +``` + +```sql +ADMIN SHOW DDL JOBS 1; +``` + +``` ++--------+---------+------------------+---------------+----------------------+-----------+----------+-----------+----------------------------+----------------------------+----------------------------+---------+-----------------------------+ +| JOB_ID | DB_NAME | TABLE_NAME | JOB_TYPE | SCHEMA_STATE | SCHEMA_ID | TABLE_ID | ROW_COUNT | CREATE_TIME | START_TIME | END_TIME | STATE | COMMENTS | ++--------+---------+------------------+---------------+----------------------+-----------+----------+-----------+----------------------------+----------------------------+----------------------------+---------+-----------------------------+ +| 153 | test | s | modify column | write reorganization | 2 | 148 | 12582912 | 2025-10-29 00:26:49.240000 | 2025-10-29 00:26:49.244000 | NULL | running | analyzing | ++--------+---------+------------------+---------------+----------------------+-----------+----------+-----------+----------------------------+----------------------------+----------------------------+---------+-----------------------------+ +1 rows in set (0.001 sec) +``` + +From the `MODIFY COLUMN` example, when `tidb_stats_update_during_ddl` is `ON`, you can see that after the execution of the `MODIFY COLUMN` DDL statement, the subsequent `EXPLAIN` output shows that statistics for the index `idx` have been automatically collected and loaded into memory (you can verify it by executing `SHOW STATS_HISTOGRAMS`). As a result, the optimizer can immediately use these statistics for range scans. If index creation or reorganization and `ANALYZE` take a long time, you can check the DDL job status by executing `ADMIN SHOW DDL JOBS`. When the `COMMENTS` column in the output contains `analyzing`, it means that the DDL job is collecting statistics. diff --git a/system-variables.md b/system-variables.md index daf752224e3cd..c45a3b575e081 100644 --- a/system-variables.md +++ b/system-variables.md @@ -1655,6 +1655,14 @@ mysql> SELECT job_info FROM mysql.analyze_jobs ORDER BY end_time DESC LIMIT 1; +### tidb_stats_update_during_ddl New in v8.5.4 and v9.0.0 + +- Scope: GLOBAL +- Persists to cluster: Yes +- Applies to hint [SET_VAR](/optimizer-hints.md#set_varvar_namevar_value): No +- Default value: `OFF` +- This variable controls whether to enable DDL-embedded `ANALYZE`. When enabled, DDL statements that create new indexes ([`ADD INDEX`](/sql-statements/sql-statement-add-index.md)) or reorganize existing indexes ([`MODIFY COLUMN`](/sql-statements/sql-statement-modify-column.md) and [`CHANGE COLUMN`](/sql-statements/sql-statement-change-column.md)) automatically collect statistics before the index becomes visible. For more information, see [`ANALYZE` Embedded in DDL Statements](/ddl_embedded_analyze.md). + ### tidb_enable_dist_task New in v7.1.0 - Scope: GLOBAL