用 EXPLAIN 查看子查询的执行计划

TiDB 会执行多种子查询相关的优化,以提升子查询的执行性能。本文档介绍一些常见子查询的优化方式,以及如何解读 EXPLAIN 语句返回的执行计划信息。

本文档所使用的示例表数据如下:

CREATE TABLE t1 (id BIGINT NOT NULL PRIMARY KEY auto_increment, pad1 BLOB, pad2 BLOB, pad3 BLOB, int_col INT NOT NULL DEFAULT 0); CREATE TABLE t2 (id BIGINT NOT NULL PRIMARY KEY auto_increment, t1_id BIGINT NOT NULL, pad1 BLOB, pad2 BLOB, pad3 BLOB, INDEX(t1_id)); CREATE TABLE t3 ( id INT NOT NULL PRIMARY KEY auto_increment, t1_id INT NOT NULL, UNIQUE (t1_id) ); INSERT INTO t1 SELECT NULL, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024), 0 FROM dual; INSERT INTO t1 SELECT NULL, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024), 0 FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t1 SELECT NULL, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024), 0 FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t1 SELECT NULL, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024), 0 FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t1 SELECT NULL, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024), 0 FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t1 SELECT NULL, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024), 0 FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t1 SELECT NULL, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024), 0 FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t1 SELECT NULL, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024), 0 FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t1 SELECT NULL, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024), 0 FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t1 SELECT NULL, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024), 0 FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t1 SELECT NULL, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024), 0 FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t2 SELECT NULL, a.id, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t2 SELECT NULL, a.id, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t2 SELECT NULL, a.id, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t2 SELECT NULL, a.id, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t2 SELECT NULL, a.id, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t2 SELECT NULL, a.id, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t2 SELECT NULL, a.id, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t2 SELECT NULL, a.id, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; INSERT INTO t2 SELECT NULL, a.id, RANDOM_BYTES(1024), RANDOM_BYTES(1024), RANDOM_BYTES(1024) FROM t1 a JOIN t1 b JOIN t1 c LIMIT 10000; UPDATE t1 SET int_col = 1 WHERE pad1 = (SELECT pad1 FROM t1 ORDER BY RAND() LIMIT 1); INSERT INTO t3 SELECT NULL, id FROM t1 WHERE id < 1000; SELECT SLEEP(1); ANALYZE TABLE t1, t2, t3;

Inner join(无 UNIQUE 约束的子查询)

以下示例中,IN 子查询会从表 t2 中搜索一列 ID。为保证语义正确性,TiDB 需要保证 t1_id 列的值具有唯一性。使用 EXPLAIN 可查看到该查询的执行计划去掉重复项并执行 Inner Join 内连接操作:

EXPLAIN SELECT * FROM t1 WHERE id IN (SELECT t1_id FROM t2);
+--------------------------------+----------+-----------+------------------------------+---------------------------------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +--------------------------------+----------+-----------+------------------------------+---------------------------------------------------------------------------------------------------------------------------+ | IndexJoin_15 | 21.11 | root | | inner join, inner:TableReader_12, outer key:test.t2.t1_id, inner key:test.t1.id, equal cond:eq(test.t2.t1_id, test.t1.id) | | ├─StreamAgg_44(Build) | 21.11 | root | | group by:test.t2.t1_id, funcs:firstrow(test.t2.t1_id)->test.t2.t1_id | | │ └─IndexReader_45 | 21.11 | root | | index:StreamAgg_34 | | │ └─StreamAgg_34 | 21.11 | cop[tikv] | | group by:test.t2.t1_id, | | │ └─IndexFullScan_26 | 90000.00 | cop[tikv] | table:t2, index:t1_id(t1_id) | keep order:true | | └─TableReader_12(Probe) | 21.11 | root | | data:TableRangeScan_11 | | └─TableRangeScan_11 | 21.11 | cop[tikv] | table:t1 | range: decided by [test.t2.t1_id], keep order:false | +--------------------------------+----------+-----------+------------------------------+---------------------------------------------------------------------------------------------------------------------------+

由上述查询结果可知,TiDB 通过索引连接操作 | IndexJoin_14 将子查询做了连接转化。该执行计划首先在 TiKV 侧通过索引扫描算子 └─IndexFullScan_31 读取 t2.t1_id 列的值,然后由 └─StreamAgg_39 算子的部分任务在 TiKV 中对 t1_id 值进行去重,然后采用 ├─StreamAgg_49(Build) 算子的部分任务在 TiDB 中对 t1_id 值再次进行去重,去重操作由聚合函数 firstrow(test.t2.t1_id) 执行;之后将操作结果与 t1 表的主键相连接,连接条件是 eq(test.t1.id, test.t2.t1_id)

Inner join(有 UNIQUE 约束的子查询)

在上述示例中,为了确保 t1_id 值在与表 t1 连接前具有唯一性,需要执行聚合运算。在以下示例中,由于 UNIQUE 约束已能确保 t3.t1_id 列值的唯一:

EXPLAIN SELECT * FROM t1 WHERE id IN (SELECT t1_id FROM t3);
+-----------------------------+---------+-----------+------------------------------+---------------------------------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +-----------------------------+---------+-----------+------------------------------+---------------------------------------------------------------------------------------------------------------------------+ | IndexJoin_18 | 999.00 | root | | inner join, inner:TableReader_15, outer key:test.t3.t1_id, inner key:test.t1.id, equal cond:eq(test.t3.t1_id, test.t1.id) | | ├─IndexReader_41(Build) | 999.00 | root | | index:IndexFullScan_40 | | │ └─IndexFullScan_40 | 999.00 | cop[tikv] | table:t3, index:t1_id(t1_id) | keep order:false | | └─TableReader_15(Probe) | 999.00 | root | | data:TableRangeScan_14 | | └─TableRangeScan_14 | 999.00 | cop[tikv] | table:t1 | range: decided by [test.t3.t1_id], keep order:false | +-----------------------------+---------+-----------+------------------------------+---------------------------------------------------------------------------------------------------------------------------+

从语义上看,因为约束保证了 t3.t1_id 列值的唯一性,TiDB 可以直接执行 INNER JOIN 查询。

Semi Join(关联查询)

在前两个示例中,通过 StreamAgg 聚合操作或通过 UNIQUE 约束保证子查询数据的唯一性之后,TiDB 才能够执行 Inner Join 操作。这两种连接均使用了 Index Join

下面的例子中,TiDB 优化器则选择了一种不同的执行计划:

EXPLAIN SELECT * FROM t1 WHERE id IN (SELECT t1_id FROM t2 WHERE t1_id != t1.int_col);
+-----------------------------+----------+-----------+------------------------------+--------------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +-----------------------------+----------+-----------+------------------------------+--------------------------------------------------------------------------------------------------------+ | MergeJoin_9 | 45446.40 | root | | semi join, left key:test.t1.id, right key:test.t2.t1_id, other cond:ne(test.t2.t1_id, test.t1.int_col) | | ├─IndexReader_24(Build) | 90000.00 | root | | index:IndexFullScan_23 | | │ └─IndexFullScan_23 | 90000.00 | cop[tikv] | table:t2, index:t1_id(t1_id) | keep order:true | | └─TableReader_22(Probe) | 56808.00 | root | | data:Selection_21 | | └─Selection_21 | 56808.00 | cop[tikv] | | ne(test.t1.id, test.t1.int_col) | | └─TableFullScan_20 | 71010.00 | cop[tikv] | table:t1 | keep order:true | +-----------------------------+----------+-----------+------------------------------+--------------------------------------------------------------------------------------------------------+

由上述查询结果可知,TiDB 执行了 Semi Join。不同于 Inner JoinSemi Join 仅允许右键 (t2.t1_id) 上的第一个值,也就是该操作将去除 Join 算子任务中的重复数据。Join 算法也包含 Merge Join,会按照排序顺序同时从左侧和右侧读取数据,这是一种高效的 Zipper Merge

可以将原语句视为关联子查询,因为它引入了子查询外的 t1.int_col 列。然而,EXPLAIN 语句的返回结果显示的是关联子查询去关联后的执行计划。条件 t1_id != t1.int_col 会被重写为 t1.id != t1.int_col。TiDB 可以从表 t1 中读取数据并且在 └─Selection_21 中执行此操作,因此这种去关联和重写操作会极大提高执行效率。

Anti Semi Join(NOT IN 子查询)

在以下示例中,除非子查询中存在 t3.t1_id,否则该查询将(从语义上)返回表 t3 中的所有行:

EXPLAIN SELECT * FROM t3 WHERE t1_id NOT IN (SELECT id FROM t1 WHERE int_col < 100);
+-----------------------------+---------+-----------+---------------+-------------------------------------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +-----------------------------+---------+-----------+---------------+-------------------------------------------------------------------------------------------------------------------------------+ | IndexJoin_16 | 799.20 | root | | anti semi join, inner:TableReader_12, outer key:test.t3.t1_id, inner key:test.t1.id, equal cond:eq(test.t3.t1_id, test.t1.id) | | ├─TableReader_28(Build) | 999.00 | root | | data:TableFullScan_27 | | │ └─TableFullScan_27 | 999.00 | cop[tikv] | table:t3 | keep order:false | | └─TableReader_12(Probe) | 999.00 | root | | data:Selection_11 | | └─Selection_11 | 999.00 | cop[tikv] | | lt(test.t1.int_col, 100) | | └─TableRangeScan_10 | 999.00 | cop[tikv] | table:t1 | range: decided by [test.t3.t1_id], keep order:false | +-----------------------------+---------+-----------+---------------+-------------------------------------------------------------------------------------------------------------------------------+

上述查询首先读取了表 t3,然后根据主键开始探测 (probe) 表 t1。连接类型是 anti semi join,即反半连接:之所以使用 anti,是因为上述示例有不存在匹配值(即 NOT IN)的情况;使用 Semi Join 则是因为仅需要匹配第一行后就可以停止查询。

Null-Aware Semi Join(IN= ANY 子查询)

IN= ANY 的集合运算符号具有特殊的三值属性(truefalseNULL)。这意味着在该运算符所转化得到的 Join 类型中需要对 Join key 两侧的 NULL 进行特殊的感知和处理。

IN= ANY 算子引导的子查询会分别转为 Semi Join 和 Left Outer Semi Join。在上述 Semi Join 小节中,示例中 Join key 两侧的列 test.t1.idtest.t2.t1_id 都为 not NULL 属性,所以 Semi Join 本身不需要 Null-Aware 的性质来辅助运算,即不需要特殊处理 NULL。当前 TiDB 对于 Null-Aware Semi Join 没有特定的优化,其实现本质都是基于笛卡尔积加过滤 (filter) 的模式。以下为 Null-Aware Semi Join 的例子:

CREATE TABLE t(a INT, b INT); CREATE TABLE s(a INT, b INT); EXPLAIN SELECT (a,b) IN (SELECT * FROM s) FROM t; EXPLAIN SELECT * FROM t WHERE (a,b) IN (SELECT * FROM s);
tidb> EXPLAIN SELECT (a,b) IN (SELECT * FROM s) FROM t; +-----------------------------+---------+-----------+---------------+-------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +-----------------------------+---------+-----------+---------------+-------------------------------------------------------------------------------------------+ | HashJoin_8 | 1.00 | root | | CARTESIAN left outer semi join, other cond:eq(test.t.a, test.s.a), eq(test.t.b, test.s.b) | | ├─TableReader_12(Build) | 1.00 | root | | data:TableFullScan_11 | | │ └─TableFullScan_11 | 1.00 | cop[tikv] | table:s | keep order:false, stats:pseudo | | └─TableReader_10(Probe) | 1.00 | root | | data:TableFullScan_9 | | └─TableFullScan_9 | 1.00 | cop[tikv] | table:t | keep order:false, stats:pseudo | +-----------------------------+---------+-----------+---------------+-------------------------------------------------------------------------------------------+ 5 rows in set (0.00 sec) tidb> EXPLAIN SELECT * FROM t WHERE (a,b) IN (SELECT * FROM s); +------------------------------+---------+-----------+---------------+-----------------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +------------------------------+---------+-----------+---------------+-----------------------------------------------------------------------------------------------------+ | HashJoin_11 | 1.00 | root | | inner join, equal:[eq(test.t.a, test.s.a) eq(test.t.b, test.s.b)] | | ├─TableReader_14(Build) | 1.00 | root | | data:Selection_13 | | │ └─Selection_13 | 1.00 | cop[tikv] | | not(isnull(test.t.a)), not(isnull(test.t.b)) | | │ └─TableFullScan_12 | 1.00 | cop[tikv] | table:t | keep order:false, stats:pseudo | | └─HashAgg_17(Probe) | 1.00 | root | | group by:test.s.a, test.s.b, funcs:firstrow(test.s.a)->test.s.a, funcs:firstrow(test.s.b)->test.s.b | | └─TableReader_24 | 1.00 | root | | data:Selection_23 | | └─Selection_23 | 1.00 | cop[tikv] | | not(isnull(test.s.a)), not(isnull(test.s.b)) | | └─TableFullScan_22 | 1.00 | cop[tikv] | table:s | keep order:false, stats:pseudo | +------------------------------+---------+-----------+---------------+-----------------------------------------------------------------------------------------------------+ 8 rows in set (0.01 sec)

第一个查询 EXPLAIN SELECT (a,b) IN (SELECT * FROM s) FROM t; 中,由于 t 表和 s 表的 ab 列都是 NULLABLE 的,所以 IN 子查询所转化的 Left Outer Semi Join 是具有 Null-Aware 性质的。具体实现是先进行笛卡尔积,然后将 IN= ANY 所连接的列作为普通等值条件放到 other condition 进行过滤(filter)。

第二个查询 EXPLAIN SELECT * FROM t WHERE (a,b) IN (SELECT * FROM s); 中,由于 t 表和 s 表的 ab 列都是 NULLABLE 的,IN 子查询本应该转为具有 Null-Aware 性质的 Semi Join,但当前 TiDB 进行了优化,直接将 Semi Join 转为了 Inner Join + Aggregate 的方式来实现。这是因为在非 scalar 输出的 IN 子查询中,NULLfalse 是等效的。下推过滤的 NULL 行导致了 WHERE 子句的否定语义,因此可以事先忽略这些行。

Null-Aware Anti Semi Join(NOT IN!= ALL 子查询)

NOT IN!= ALL 的集合运算运算具有特殊的三值属性(truefalseNULL)。这意味着在其所转化得到的 Join 类型中需要对 Join key 两侧的 NULL 进行特殊的感知和处理。

NOT IN!= ALL 算子引导的子查询会对应地转为 Anti Semi Join 和 Anti Left Outer Semi Join。在上述的 Anti Semi Join 小节中,由于示例中 Join key 两侧的列 test.t3.t1_idtest.t1.id 都是 not NULL 属性的,所以 Anti Semi Join 本身不需要 Null-Aware 的性质来辅助计算,即不需要特殊处理 NULL

在 TiDB v6.3.0 版本,TiDB 引入了针对 Null-Aware Anti Join (NAAJ) 的如下特殊优化:

  • 利用 Null-Aware 的等值条件 (NA-EQ) 构建哈希连接

    由于集合操作符引入的等值需要对等值两侧操作符数的 NULL 值做特殊处理,这里称需要 Null-Aware 的等值条件为 NA-EQ 条件。与 v6.3.0 之前版本不同的是,TiDB 不会再将 NA-EQ 条件处理成普通 EQ 条件,而是专门放置于 Join 后置的 other condition 中,匹配笛卡尔积后再判断结果集的合法性。

    在 TiDB v6.3.0 版本中,NA-EQ 这种弱化的等值条件依然会被用来构建哈希值 (Hash Join),大大减少了匹配时所需遍历的数据量,加速匹配过程。在 build 表 DISTINCT 值比例趋近 1 的时候,加速效果更为显著。

  • 利用两侧数据源 NULL 值的特殊性质加速匹配过程的返回

    由于 Anti Semi Join 自身具有 CNF (Conjunctive normal form) 表达式的属性,其任何一侧出现的 NULL 值都会导致确定的结果。利用这个性质可以来加速整个匹配过程。

以下为 Null-Aware Anti Semi Join 的例子:

CREATE TABLE t(a INT, b INT); CREATE TABLE s(a INT, b INT); EXPLAIN SELECT (a, b) NOT IN (SELECT * FROM s) FROM t; EXPLAIN SELECT * FROM t WHERE (a, b) NOT IN (SELECT * FROM s);
tidb> EXPLAIN SELECT (a, b) NOT IN (SELECT * FROM s) FROM t; +-----------------------------+----------+-----------+---------------+---------------------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +-----------------------------+----------+-----------+---------------+---------------------------------------------------------------------------------------------+ | HashJoin_8 | 10000.00 | root | | Null-aware anti left outer semi join, equal:[eq(test.t.b, test.s.b) eq(test.t.a, test.s.a)] | | ├─TableReader_12(Build) | 10000.00 | root | | data:TableFullScan_11 | | │ └─TableFullScan_11 | 10000.00 | cop[tikv] | table:s | keep order:false, stats:pseudo | | └─TableReader_10(Probe) | 10000.00 | root | | data:TableFullScan_9 | | └─TableFullScan_9 | 10000.00 | cop[tikv] | table:t | keep order:false, stats:pseudo | +-----------------------------+----------+-----------+---------------+---------------------------------------------------------------------------------------------+ 5 rows in set (0.00 sec) tidb> EXPLAIN SELECT * FROM t WHERE (a, b) NOT IN (SELECT * FROM s); +-----------------------------+----------+-----------+---------------+----------------------------------------------------------------------------------+ | id | estRows | task | access object | operator info | +-----------------------------+----------+-----------+---------------+----------------------------------------------------------------------------------+ | HashJoin_8 | 8000.00 | root | | Null-aware anti semi join, equal:[eq(test.t.b, test.s.b) eq(test.t.a, test.s.a)] | | ├─TableReader_12(Build) | 10000.00 | root | | data:TableFullScan_11 | | │ └─TableFullScan_11 | 10000.00 | cop[tikv] | table:s | keep order:false, stats:pseudo | | └─TableReader_10(Probe) | 10000.00 | root | | data:TableFullScan_9 | | └─TableFullScan_9 | 10000.00 | cop[tikv] | table:t | keep order:false, stats:pseudo | +-----------------------------+----------+-----------+---------------+----------------------------------------------------------------------------------+ 5 rows in set (0.00 sec)

第一个查询 EXPLAIN SELECT (a, b) NOT IN (SELECT * FROM s) FROM t; 中,由于 t 表和 s 表的 ab 列都是 NULLABLE 的,所以 NOT IN 子查询所转化的 Left Outer Semi Join 是具有 Null-Aware 性质的。不同的是,NAAJ 优化将 NA-EQ 条件也作为了 Hash Join 的连接条件,大大加速了 Join 的计算。

第二个查询 EXPLAIN SELECT * FROM t WHERE (a, b) NOT IN (SELECT * FROM s); 中,由于 t 表和 s 表的 ab 列都是 NULLABLE 的,所以 NOT IN 子查询所转化的 Anti Semi Join 是具有 Null-Aware 性质的。不同的是,NAAJ 优化将 NA-EQ 条件也作为了 Hash Join 的连接条件,大大加速了 Join 的计算。

当前 TiDB 仅针对 Anti Semi Join 和 Anti Left Outer Semi Join 实现了 NULL 感知。目前仅支持 Hash Join 类型且其 build 表只能固定为右侧表。

其他类型查询的执行计划