用 EXPLAIN 查看索引查询的执行计划

SQL 查询可能会使用索引,可以通过 EXPLAIN 语句来查看索引查询的执行计划。本文提供多个示例,以帮助用户理解索引查询是如何执行的。

TiDB 支持以下使用索引的算子来提升查询速度:

本文档中的示例都基于以下数据:

CREATE TABLE t1 ( id INT NOT NULL PRIMARY KEY auto_increment, intkey INT NOT NULL, pad1 VARBINARY(1024), INDEX (intkey) ); INSERT INTO t1 SELECT NULL, FLOOR(RAND()*1024), RANDOM_BYTES(1024) FROM dual; INSERT INTO t1 SELECT NULL, FLOOR(RAND()*1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t1 SELECT NULL, FLOOR(RAND()*1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t1 SELECT NULL, FLOOR(RAND()*1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000;

IndexLookup

TiDB 从二级索引检索数据时会使用 IndexLookup 算子。例如,以下所有查询均会在 intkey 列的索引上使用 IndexLookup 算子:

EXPLAIN SELECT * FROM t1 WHERE intkey = 123; EXPLAIN SELECT * FROM t1 WHERE intkey < 10; EXPLAIN SELECT * FROM t1 WHERE intkey BETWEEN 300 AND 310; EXPLAIN SELECT * FROM t1 WHERE intkey IN (123,29,98); EXPLAIN SELECT * FROM t1 WHERE intkey >= 99 AND intkey <= 103;
+-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ | id | estRows | task | access object | operator info | +-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ | IndexLookUp_10 | 1.00 | root | | | | ├─IndexRangeScan_8(Build) | 1.00 | cop[tikv] | table:t1, index:intkey(intkey) | range:[123,123], keep order:false | | └─TableRowIDScan_9(Probe) | 1.00 | cop[tikv] | table:t1 | keep order:false | +-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ 3 rows in set (0.00 sec) +-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ | id | estRows | task | access object | operator info | +-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ | IndexLookUp_10 | 3.60 | root | | | | ├─IndexRangeScan_8(Build) | 3.60 | cop[tikv] | table:t1, index:intkey(intkey) | range:[-inf,10), keep order:false | | └─TableRowIDScan_9(Probe) | 3.60 | cop[tikv] | table:t1 | keep order:false | +-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ 3 rows in set (0.00 sec) +-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ | id | estRows | task | access object | operator info | +-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ | IndexLookUp_10 | 5.67 | root | | | | ├─IndexRangeScan_8(Build) | 5.67 | cop[tikv] | table:t1, index:intkey(intkey) | range:[300,310], keep order:false | | └─TableRowIDScan_9(Probe) | 5.67 | cop[tikv] | table:t1 | keep order:false | +-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ 3 rows in set (0.00 sec) +-------------------------------+---------+-----------+--------------------------------+-----------------------------------------------------+ | id | estRows | task | access object | operator info | +-------------------------------+---------+-----------+--------------------------------+-----------------------------------------------------+ | IndexLookUp_10 | 4.00 | root | | | | ├─IndexRangeScan_8(Build) | 4.00 | cop[tikv] | table:t1, index:intkey(intkey) | range:[29,29], [98,98], [123,123], keep order:false | | └─TableRowIDScan_9(Probe) | 4.00 | cop[tikv] | table:t1 | keep order:false | +-------------------------------+---------+-----------+--------------------------------+-----------------------------------------------------+ 3 rows in set (0.00 sec) +-------------------------------+---------+-----------+--------------------------------+----------------------------------+ | id | estRows | task | access object | operator info | +-------------------------------+---------+-----------+--------------------------------+----------------------------------+ | IndexLookUp_10 | 6.00 | root | | | | ├─IndexRangeScan_8(Build) | 6.00 | cop[tikv] | table:t1, index:intkey(intkey) | range:[99,103], keep order:false | | └─TableRowIDScan_9(Probe) | 6.00 | cop[tikv] | table:t1 | keep order:false | +-------------------------------+---------+-----------+--------------------------------+----------------------------------+ 3 rows in set (0.00 sec)

IndexLookup 算子有以下两个子节点:

  • ├─IndexRangeScan_8(Build) 算子节点对 intkey 列的索引执行范围扫描,并检索内部的 RowID 值(对此表而言,即为主键)。
  • └─TableRowIDScan_9(Probe) 算子节点随后从表数据中检索整行。

IndexLookup 任务分以上两步执行。如果满足条件的行较多,SQL 优化器可能会根据统计信息选择使用 TableFullScan 算子。在以下示例中,很多行都满足 intkey > 100 这一条件,因此优化器选择了 TableFullScan

EXPLAIN SELECT * FROM t1 WHERE intkey > 100;
+-------------------------+---------+-----------+---------------+-------------------------+ | id | estRows | task | access object | operator info | +-------------------------+---------+-----------+---------------+-------------------------+ | TableReader_7 | 898.50 | root | | data:Selection_6 | | └─Selection_6 | 898.50 | cop[tikv] | | gt(test.t1.intkey, 100) | | └─TableFullScan_5 | 1010.00 | cop[tikv] | table:t1 | keep order:false | +-------------------------+---------+-----------+---------------+-------------------------+ 3 rows in set (0.00 sec)

IndexLookup 算子能在带索引的列上有效优化 LIMIT

EXPLAIN SELECT * FROM t1 ORDER BY intkey DESC LIMIT 10;
+--------------------------------+---------+-----------+--------------------------------+------------------------------------+ | id | estRows | task | access object | operator info | +--------------------------------+---------+-----------+--------------------------------+------------------------------------+ | IndexLookUp_21 | 10.00 | root | | limit embedded(offset:0, count:10) | | ├─Limit_20(Build) | 10.00 | cop[tikv] | | offset:0, count:10 | | │ └─IndexFullScan_18 | 10.00 | cop[tikv] | table:t1, index:intkey(intkey) | keep order:true, desc | | └─TableRowIDScan_19(Probe) | 10.00 | cop[tikv] | table:t1 | keep order:false, stats:pseudo | +--------------------------------+---------+-----------+--------------------------------+------------------------------------+ 4 rows in set (0.00 sec)

以上示例中,TiDB 从 intkey 索引读取最后 10 行,然后从表数据中检索这些行的 RowID 值。

IndexReader

TiDB 支持覆盖索引优化 (covering index optimization)。如果 TiDB 能从索引中检索出所有行,就会跳过 IndexLookup 任务中通常所需的第二步(即从表数据中检索整行)。示例如下:

EXPLAIN SELECT * FROM t1 WHERE intkey = 123; EXPLAIN SELECT id FROM t1 WHERE intkey = 123;
+-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ | id | estRows | task | access object | operator info | +-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ | IndexLookUp_10 | 1.00 | root | | | | ├─IndexRangeScan_8(Build) | 1.00 | cop[tikv] | table:t1, index:intkey(intkey) | range:[123,123], keep order:false | | └─TableRowIDScan_9(Probe) | 1.00 | cop[tikv] | table:t1 | keep order:false | +-------------------------------+---------+-----------+--------------------------------+-----------------------------------+ 3 rows in set (0.00 sec) +--------------------------+---------+-----------+--------------------------------+-----------------------------------+ | id | estRows | task | access object | operator info | +--------------------------+---------+-----------+--------------------------------+-----------------------------------+ | Projection_4 | 1.00 | root | | test.t1.id | | └─IndexReader_6 | 1.00 | root | | index:IndexRangeScan_5 | | └─IndexRangeScan_5 | 1.00 | cop[tikv] | table:t1, index:intkey(intkey) | range:[123,123], keep order:false | +--------------------------+---------+-----------+--------------------------------+-----------------------------------+ 3 rows in set (0.00 sec)

以上结果中,id 也是内部的 RowID 值,因此 id 也存储在 intkey 索引中。部分 └─IndexRangeScan_5 任务使用 intkey 索引后,可直接返回 RowID 值。

Point_Get 和 Batch_Point_Get

TiDB 直接从主键或唯一键检索数据时会使用 Point_GetBatch_Point_Get 算子。这两个算子比 IndexLookup 更有效率。示例如下:

EXPLAIN SELECT * FROM t1 WHERE id = 1234; EXPLAIN SELECT * FROM t1 WHERE id IN (1234,123); ALTER TABLE t1 ADD unique_key INT; UPDATE t1 SET unique_key = id; ALTER TABLE t1 ADD UNIQUE KEY (unique_key); EXPLAIN SELECT * FROM t1 WHERE unique_key = 1234; EXPLAIN SELECT * FROM t1 WHERE unique_key IN (1234, 123);
+-------------+---------+------+---------------+---------------+ | id | estRows | task | access object | operator info | +-------------+---------+------+---------------+---------------+ | Point_Get_1 | 1.00 | root | table:t1 | handle:1234 | +-------------+---------+------+---------------+---------------+ 1 row in set (0.00 sec) +-------------------+---------+------+---------------+-------------------------------------------------+ | id | estRows | task | access object | operator info | +-------------------+---------+------+---------------+-------------------------------------------------+ | Batch_Point_Get_1 | 2.00 | root | table:t1 | handle:[1234 123], keep order:false, desc:false | +-------------------+---------+------+---------------+-------------------------------------------------+ 1 row in set (0.00 sec) Query OK, 0 rows affected (0.27 sec) Query OK, 1010 rows affected (0.06 sec) Rows matched: 1010 Changed: 1010 Warnings: 0 Query OK, 0 rows affected (0.37 sec) +-------------+---------+------+----------------------------------------+---------------+ | id | estRows | task | access object | operator info | +-------------+---------+------+----------------------------------------+---------------+ | Point_Get_1 | 1.00 | root | table:t1, index:unique_key(unique_key) | | +-------------+---------+------+----------------------------------------+---------------+ 1 row in set (0.00 sec) +-------------------+---------+------+----------------------------------------+------------------------------+ | id | estRows | task | access object | operator info | +-------------------+---------+------+----------------------------------------+------------------------------+ | Batch_Point_Get_1 | 2.00 | root | table:t1, index:unique_key(unique_key) | keep order:false, desc:false | +-------------------+---------+------+----------------------------------------+------------------------------+ 1 row in set (0.00 sec)

IndexFullScan

索引是有序的,所以优化器可以使用 IndexFullScan 算子来优化常见的查询,例如在索引值上使用 MINMax 函数:

EXPLAIN SELECT MIN(intkey) FROM t1; EXPLAIN SELECT MAX(intkey) FROM t1;
+------------------------------+---------+-----------+--------------------------------+-------------------------------------+ | id | estRows | task | access object | operator info | +------------------------------+---------+-----------+--------------------------------+-------------------------------------+ | StreamAgg_12 | 1.00 | root | | funcs:min(test.t1.intkey)->Column#4 | | └─Limit_16 | 1.00 | root | | offset:0, count:1 | | └─IndexReader_29 | 1.00 | root | | index:Limit_28 | | └─Limit_28 | 1.00 | cop[tikv] | | offset:0, count:1 | | └─IndexFullScan_27 | 1.00 | cop[tikv] | table:t1, index:intkey(intkey) | keep order:true | +------------------------------+---------+-----------+--------------------------------+-------------------------------------+ 5 rows in set (0.00 sec) +------------------------------+---------+-----------+--------------------------------+-------------------------------------+ | id | estRows | task | access object | operator info | +------------------------------+---------+-----------+--------------------------------+-------------------------------------+ | StreamAgg_12 | 1.00 | root | | funcs:max(test.t1.intkey)->Column#4 | | └─Limit_16 | 1.00 | root | | offset:0, count:1 | | └─IndexReader_29 | 1.00 | root | | index:Limit_28 | | └─Limit_28 | 1.00 | cop[tikv] | | offset:0, count:1 | | └─IndexFullScan_27 | 1.00 | cop[tikv] | table:t1, index:intkey(intkey) | keep order:true, desc | +------------------------------+---------+-----------+--------------------------------+-------------------------------------+ 5 rows in set (0.00 sec)

以上语句的执行过程中,TiDB 在每一个 TiKV Region 上执行 IndexFullScan 操作。虽然算子名为 FullScan 即全扫描,TiDB 只读取第一行 (└─Limit_28)。每个 TiKV Region 返回各自的 MINMAX 值给 TiDB,TiDB 再执行流聚合运算来过滤出一行数据。即使表为空,带 MAXMIN 函数的流聚合运算也能保证返回 NULL 值。

相反,在没有索引的值上执行 MIN 函数会在每一个 TiKV Region 上执行 TableFullScan 操作。该查询会要求在 TiKV 中扫描所有行,但 TopN 计算可保证每个 TiKV Region 只返回一行数据给 TiDB。尽管 TopN 能减少 TiDB 和 TiKV 之间的多余数据传输,但该查询的效率仍远不及以上示例(MIN 能够使用索引)。

EXPLAIN SELECT MIN(pad1) FROM t1;
+--------------------------------+---------+-----------+---------------+-----------------------------------+ | id | estRows | task | access object | operator info | +--------------------------------+---------+-----------+---------------+-----------------------------------+ | StreamAgg_13 | 1.00 | root | | funcs:min(test.t1.pad1)->Column#4 | | └─TopN_14 | 1.00 | root | | test.t1.pad1, offset:0, count:1 | | └─TableReader_23 | 1.00 | root | | data:TopN_22 | | └─TopN_22 | 1.00 | cop[tikv] | | test.t1.pad1, offset:0, count:1 | | └─Selection_21 | 1008.99 | cop[tikv] | | not(isnull(test.t1.pad1)) | | └─TableFullScan_20 | 1010.00 | cop[tikv] | table:t1 | keep order:false | +--------------------------------+---------+-----------+---------------+-----------------------------------+ 6 rows in set (0.00 sec)

执行以下语句时,TiDB 将使用 IndexFullScan 算子扫描索引中的每一行:

EXPLAIN SELECT SUM(intkey) FROM t1; EXPLAIN SELECT AVG(intkey) FROM t1;
+----------------------------+---------+-----------+--------------------------------+-------------------------------------+ | id | estRows | task | access object | operator info | +----------------------------+---------+-----------+--------------------------------+-------------------------------------+ | StreamAgg_20 | 1.00 | root | | funcs:sum(Column#6)->Column#4 | | └─IndexReader_21 | 1.00 | root | | index:StreamAgg_8 | | └─StreamAgg_8 | 1.00 | cop[tikv] | | funcs:sum(test.t1.intkey)->Column#6 | | └─IndexFullScan_19 | 1010.00 | cop[tikv] | table:t1, index:intkey(intkey) | keep order:false | +----------------------------+---------+-----------+--------------------------------+-------------------------------------+ 4 rows in set (0.00 sec) +----------------------------+---------+-----------+--------------------------------+----------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +----------------------------+---------+-----------+--------------------------------+----------------------------------------------------------------------------+ | StreamAgg_20 | 1.00 | root | | funcs:avg(Column#7, Column#8)->Column#4 | | └─IndexReader_21 | 1.00 | root | | index:StreamAgg_8 | | └─StreamAgg_8 | 1.00 | cop[tikv] | | funcs:count(test.t1.intkey)->Column#7, funcs:sum(test.t1.intkey)->Column#8 | | └─IndexFullScan_19 | 1010.00 | cop[tikv] | table:t1, index:intkey(intkey) | keep order:false | +----------------------------+---------+-----------+--------------------------------+----------------------------------------------------------------------------+ 4 rows in set (0.00 sec)

以上示例中,IndexFullScanTableFullScan 更有效率,因为 (intkey + RowID) 索引中值的长度小于整行的长度。

以下语句不支持使用 IndexFullScan 算子,因为涉及该表中的其他列:

EXPLAIN SELECT AVG(intkey), ANY_VALUE(pad1) FROM t1;
+------------------------------+---------+-----------+---------------+-----------------------------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +------------------------------+---------+-----------+---------------+-----------------------------------------------------------------------------------------------------------------------+ | Projection_4 | 1.00 | root | | Column#4, any_value(test.t1.pad1)->Column#5 | | └─StreamAgg_16 | 1.00 | root | | funcs:avg(Column#10, Column#11)->Column#4, funcs:firstrow(Column#12)->test.t1.pad1 | | └─TableReader_17 | 1.00 | root | | data:StreamAgg_8 | | └─StreamAgg_8 | 1.00 | cop[tikv] | | funcs:count(test.t1.intkey)->Column#10, funcs:sum(test.t1.intkey)->Column#11, funcs:firstrow(test.t1.pad1)->Column#12 | | └─TableFullScan_15 | 1010.00 | cop[tikv] | table:t1 | keep order:false | +------------------------------+---------+-----------+---------------+-----------------------------------------------------------------------------------------------------------------------+ 5 rows in set (0.00 sec)

其他类型查询的执行计划

文档内容是否有帮助?