Explain Statements That Use Subqueries

TiDB performs several optimizations to improve the performance of subqueries. This document describes some of these optimizations for common subqueries and how to interpret the output of EXPLAIN.

The examples in this document are based on the following sample data:

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 (non-unique subquery)

In the following example, the IN subquery searches for a list of IDs from the table t2. For semantic correctness, TiDB needs to guarantee that the column t1_id is unique. Using EXPLAIN, you can see the execution plan used to remove duplicates and perform an INNER JOIN operation:

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 | +--------------------------------+----------+-----------+------------------------------+---------------------------------------------------------------------------------------------------------------------------+

From the query results above, you can see that TiDB uses the index join operation | IndexJoin_14 to join and transform the subquery. In the execution plan, the execution process is as follows:

  1. The index scanning operator └─IndexFullScan_31 at the TiKV side reads the values of the t2.t1_id column.
  2. Some tasks of the └─StreamAgg_39 operator deduplicate the values of t1_id in TiKV.
  3. Some tasks of the ├─StreamAgg_49(Build) operator deduplicate the values of t1_id in TiDB. The deduplication is performed by the aggregate function firstrow(test.t2.t1_id).
  4. The operation results are joined with the primary key of the t1 table. The join condition is eq(test.t1.id, test.t2.t1_id).

Inner join (unique subquery)

In the previous example, aggregation is required to ensure that the values of t1_id are unique before joining against the table t1. But in the following example, t3.t1_id is already guaranteed unique because of a UNIQUE constraint:

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 | +-----------------------------+---------+-----------+------------------------------+---------------------------------------------------------------------------------------------------------------------------+

Semantically because t3.t1_id is guaranteed unique, it can be executed directly as an INNER JOIN.

Semi join (correlated subquery)

In the previous two examples, TiDB is able to perform an INNER JOIN operation after the data inside the subquery is made unique (via StreamAgg) or guaranteed unique. Both joins are performed using an Index Join.

In this example, TiDB chooses a different execution plan:

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 | +-----------------------------+----------+-----------+------------------------------+--------------------------------------------------------------------------------------------------------+

From the result above, you can see that TiDB uses a Semi Join algorithm. Semi-join differs from inner join: semi-join only permits the first value on the right key (t2.t1_id), which means that the duplicates are eliminated as a part of the join operator task. The join algorithm is also Merge Join, which is like an efficient zipper-merge as the operator reads data from both the left and the right side in sorted order.

The original statement is considered a correlated subquery, because the subquery refers to a column (t1.int_col) that exists outside of the subquery. However, the output of EXPLAIN shows the execution plan after the subquery decorrelation optimization has been applied. The condition t1_id != t1.int_col is rewritten to t1.id != t1.int_col. TiDB can perform this in └─Selection_21 as it is reading data from the table t1, so this decorrelation and rewriting make the execution a lot more efficient.

Anti semi join (NOT IN subquery)

In the following example, the query semantically returns all rows from the table t3 unless t3.t1_id is in the subquery:

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 | +-----------------------------+---------+-----------+---------------+-------------------------------------------------------------------------------------------------------------------------------+

This query starts by reading the table t3 and then probes the table t1 based on the PRIMARY KEY. The join type is an anti semi join; anti because this example is for the non-existence of the value (NOT IN) and semi-join because only the first row needs to match before the join is rejected.

Null-aware semi join (IN and = ANY subqueries)

The value of the IN or = ANY set operator is three-valued (true, false, and NULL). For the join type converted from either of the two operators, TiDB needs to be aware of the NULL on both sides of the join key and process it in a special way.

Subqueries containing IN and = ANY operators are converted to semi join and left outer semi join respectively. In the preceding example of Semi join, since columns test.t1.id and test.t2.t1_id on both sides of the join key are not NULL, the semi join does not need to be considered as null-aware (NULL is not processed specially). TiDB processes the null-aware semi join based on the Cartesian product and filter without special optimization. The following is an example:

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)

In the first query statement EXPLAIN SELECT (a,b) IN (SELECT * FROM s) FROM t;, since columns a and b of tables t and s are NULLABLE, the left outer semi join converted by the IN subquery is null-aware. Specifically, the Cartesian product is calculated first, then the column connected by IN or = ANY is put into other conditions as a normal equality query for filtering.

In the second query statement EXPLAIN SELECT * FROM t WHERE (a,b) IN (SELECT * FROM s);, since columns a and b of tables t and s are NULLABLE, the IN subquery should have been converted to a null-aware semi join. But TiDB optimizes it by converting semi join to inner join and aggregate. This is because NULL and false are equivalent in IN subqueries for non-scalar output. The NULL rows in the push-down filter results in the negative semantics of the WHERE clause. Therefore, these rows can be ignored beforehand.

Null-aware anti semi join (NOT IN and != ALL subqueries)

The value of the NOT IN or != ALL set operator is three-valued (true, false, and NULL). For the join type converted from either of the two operators, TiDB needs to be aware of the NULL on both sides of the join key and process it in a special way.

Subqueries containing NOT IN and ! = ALL operators are converted to anti semi join and anti left outer semi join respectively. In the preceding example of Anti semi join, since columns test.t3.t1_id and test.t1.id on both sides of the join key are not NULL, the anti semi join does not need to be considered as null-aware (NULL is not processed specially).

TiDB v6.3.0 optimizes null-aware anti join (NAAJ) as follows:

  • Build hash join using the null-aware equality condition (NA-EQ)

    Set operators introduce the equality condition, which requires a special process for the NULL value of operators on both sides of the condition. The equality condition that requires null-aware is called NA-EQ. Different from earlier versions, TiDB v6.3.0 no longer processes NA-EQ as before, but places it in other conditions after join, and then determines the legitimacy of the result set after matching the Cartesian product.

    Since TiDB v6.3.0, NA-EQ, a weakened equality condition, is still used to build hash join. This reduces the matching amount of data that needs to be traversed and speeds up the matching process. The acceleration is more significant when the percentage of total DISTINCT() values of the build table is almost 100%.

  • Speed up the return of matching results using the special property of NULL

    Since anti semi join is a conjunctive normal form (CNF), a NULL on either side of the join leads to a definite result. This property can be used to speed up the return of the entire matching process.

The following is an example:

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)

In the first query statement EXPLAIN SELECT (a, b) NOT IN (SELECT * FROM s) FROM t;, since columns a and b of tables t and s are NULLABLE, the left outer semi join converted by NOT IN subquery is null-aware. The difference is that NAAJ optimization also uses the NA-EQ as the hash join condition, which greatly speeds up the join calculation.

In the second query statement EXPLAIN SELECT * FROM t WHERE (a, b) NOT IN (SELECT * FROM s);, since columns a and b of tables t and s are NULLABLE, the anti semi join converted by NOT IN subquery is null-aware. The difference is that NAAJ optimization also uses the NA-EQ as the hash join condition, which greatly speeds up the join calculation.

Currently, TiDB can only be null-aware of anti semi join and anti left outer semi join. Only the hash join type is supported and its build table should be fixed to the right table.

Explain statements using other types of subqueries