作者： 数据源的TiDB学习之路 原文来源： https://tidb.net/blog/e9563a23
导读
TiDB 针对子查询语句会执行多种 子查询相关的优化 ，以提升子查询的执行性能。半连接语句和关联子查询语句是常用的两类子查询，TiDB 优化器默认包含一些自动优化策略，同时 TiDB 也提供额外的 HINT 用于影响优化器在特定场景下可以选择更高效的执行计划。本文针对半连接及关联子查询语句在 TiDB 中的用法及优化技巧进行说明。
半连接优化
关于半连接（Semi Join）
半连接（Semi Join）是数据库中一种特殊的连接操作，它只关心一个表（通常称为外表或左表）中的记录是否在另一个表（通常称为内表或右表）中有匹配的记录，而不关心匹配记录的具体数量或内容。半连接的结果集通常只包含外表中的记录，并且这些记录在内表中至少有一个匹配项。
在SQL中，半连接可以通过多种方式实现，包括但不限于：
- 使用EXISTS子查询 ：最常见的实现方式之一。通过在外表的主查询中使用EXISTS关键字，并嵌套一个在内表上进行查找的子查询，可以判断外表中的记录是否在内表中有匹配项。
- 使用IN子查询 ：虽然IN子查询通常用于返回匹配项的具体内容，但也可以用于实现半连接的效果。需要注意的是，当内表中的匹配项很多时，IN子查询的性能可能会下降。
- 直接转换为JOIN操作 ：在某些数据库系统中，优化器可能会自动将半连接转换为更高效的JOIN操作。这通常取决于数据库系统的具体实现和查询的复杂性。
下面是一个典型的半连接SQL语句，
select * from t1 WHERE EXISTS (SELECT 1 FROM t2 WHERE t2.a = t1.a);
优化器针对 Semi Join 优化的不足
优化器针对 Semi Join 语句中表之间的关联方式可能会转换为不同的 Join 方式，如 Hash Join 或 Index Join。然而，不论是哪种 Join 方式，在某些特定的场景下都存在一定的不足之处。
- 选择 Hash Join 时 ：Semi Join 被优化为 Hash Join 的执行方式时，只能够使用子查询构建哈希表，因此在子查询比外查询结果集大时，执行速度可能会不及预期。
- 选择 Index Join 时 ：Semi Join 被优化为 Index Join 的执行方式时，只能够使用外查询作为驱动表，因此在子查询比外查询结果集小时，执行速度可能会不及预期。
TiDB semi_join_rewrite() 优化的效果
针对上述默认优化器存在的不足，TiDB 提供 semi_join_rewrite() HINT。使用 SEMI_JOIN_REWRITE() 进行改写后，优化器便可以扩大选择范围，选择更好的执行方式。（目前该 Hint 只作用于 EXISTS 子查询）
- Hash Join 的优化效果
根据上述描述，执行计划默认选择走 Hash Join 的情况下， 只能够使用子查询构建哈希表 ，因此在子查询的结果集比外查询大时，执行速度可能不及预期。我们模拟 2 张表 t1 和 t2 分别为 100 万 和 1000 万，通过执行计划可以发现，子查询中的大表 t2 被用来构建哈希表，造成语句的执行耗时 约为 7 秒 。
注：为了确保执行计划选择 Hash Join，添加 /*+ hash_join(t1,t2@sel_2)*/ HINT。
mysql> explain analyze select /*+ hash_join(t1,t2@sel_2)*/ * from t1 WHERE EXISTS (SELECT 1 FROM t2 WHERE t2.a = t1.a);
+-----------------------------+-------------+----------+-----------+-----------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+---------------------------------------------+----------+---------+
| id | estRows | actRows | task | access object | execution info | operator info | memory | disk |
+-----------------------------+-------------+----------+-----------+-----------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+---------------------------------------------+----------+---------+
| HashJoin_21 | 800000.00 | 1000000 | root | | time:7.13s, loops:979, RU:9602.429239, build_hash_table:{total:6.85s, fetch:172ms, build:6.68s}, probe:{concurrency:5, total:35.7s, max:7.13s, probe:1.35s, fetch and wait:34.3s} | semi join, equal:[eq(test.t1.a, test.t2.a)] | 801.3 MB | 0 Bytes |
| ├─IndexReader_29(Build) | 10000000.00 | 10000000 | root | | time:93ms, loops:9783, cop_task: {num: 328, max: 50.2ms, min: 523.3µs, avg: 23.7ms, p95: 41.4ms, max_proc_keys: 50144, p95_proc_keys: 50144, tot_proc: 7.26s, tot_wait: 55.2ms, copr_cache_hit_ratio: 0.02, build_task_duration: 46.6µs, max_distsql_concurrency: 15}, rpc_info:{Cop:{num_rpc:328, total_time:7.75s}} | index:IndexFullScan_28 | 6.10 MB | N/A |
| │ └─IndexFullScan_28 | 10000000.00 | 10000000 | cop[tikv] | table:t2, index:idx_t2_a(a) | tikv_task:{proc max:50ms, min:0s, avg: 21.3ms, p80:40ms, p95:40ms, iters:11061, tasks:328}, scan_detail: {total_process_keys: 9987904, total_process_keys_size: 459443584, total_keys: 9988226, get_snapshot_time: 28.7ms, rocksdb: {key_skipped_count: 9987904, block: {cache_hit_count: 16496}}}, time_detail: {total_process_time: 7.26s, total_suspend_time: 20.5ms, total_wait_time: 55.2ms, total_kv_read_wall_time: 6.95s, tikv_wall_time: 7.43s} | keep order:false | N/A | N/A |
| └─TableReader_24(Probe) | 1000000.00 | 1000000 | root | | time:14.3ms, loops:979, cop_task: {num: 28, max: 1.68ms, min: 596.2µs, avg: 1.13ms, p95: 1.5ms, max_proc_keys: 992, p95_proc_keys: 480, tot_proc: 2.22ms, tot_wait: 3.62ms, copr_cache_hit_ratio: 0.89, build_task_duration: 21.4µs, max_distsql_concurrency: 1}, rpc_info:{Cop:{num_rpc:28, total_time:30.2ms}} | data:Selection_23 | 1.53 MB | N/A |
| └─Selection_23 | 1000000.00 | 1000000 | cop[tikv] | | tikv_task:{proc max:50ms, min:0s, avg: 32.5ms, p80:50ms, p95:50ms, iters:1088, tasks:28}, scan_detail: {total_process_keys: 1696, total_process_keys_size: 79129, total_keys: 1699, get_snapshot_time: 1.17ms, rocksdb: {key_skipped_count: 1696, block: {cache_hit_count: 15}}}, time_detail: {total_process_time: 2.22ms, total_suspend_time: 15.4µs, total_wait_time: 3.62ms, tikv_wall_time: 11.8ms} | not(isnull(test.t1.a)) | N/A | N/A |
| └─TableFullScan_22 | 1000000.00 | 1000000 | cop[tikv] | table:t1 | tikv_task:{proc max:50ms, min:0s, avg: 31.4ms, p80:50ms, p95:50ms, iters:1088, tasks:28} | keep order:false | N/A | N/A |
+-----------------------------+-------------+----------+-----------+-----------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+---------------------------------------------+----------+---------+
6 rows in set (7.13 sec)
而当我们添加了 /*+ semi_join_rewrite() */ HINT 之后，执行计划虽然还是走 Hash Join，但是被哈希的表变成了小表 t1， 同时大表 t2 也做了一个分组聚合。执行耗时也是 优化到 3.6 秒，提升了 1 倍。
mysql> explain analyze select /*+ hash_join(t1,t2@sel_2)*/ * from t1 WHERE EXISTS (SELECT /*+ semi_join_rewrite() */1 FROM t2 WHERE t2.a = t1.a);
+------------------------------+-------------+----------+-----------+-----------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+----------------------------------------------------------+----------+---------+
| id | estRows | actRows | task | access object | execution info | operator info | memory | disk |
+------------------------------+-------------+----------+-----------+-----------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+----------------------------------------------------------+----------+---------+
| HashJoin_28 | 1000000.00 | 1000000 | root | | time:3.6s, loops:980, RU:9856.929542, build_hash_table:{total:568.5ms, fetch:19.1ms, build:549.5ms}, probe:{concurrency:5, total:18s, max:3.6s, probe:4.96s, fetch and wait:13s} | inner join, equal:[eq(test.t1.a, test.t2.a)] | 102.3 MB | 0 Bytes |
| ├─TableReader_31(Build) | 1000000.00 | 1000000 | root | | time:13.2ms, loops:979, cop_task: {num: 28, max: 16ms, min: 570.1µs, avg: 1.55ms, p95: 2.28ms, max_proc_keys: 992, p95_proc_keys: 480, tot_proc: 2.43ms, tot_wait: 2.98ms, copr_cache_hit_ratio: 0.89, build_task_duration: 21µs, max_distsql_concurrency: 1}, rpc_info:{Cop:{num_rpc:28, total_time:42.4ms}} | data:Selection_30 | 1.53 MB | N/A |
| │ └─Selection_30 | 1000000.00 | 1000000 | cop[tikv] | | tikv_task:{proc max:50ms, min:0s, avg: 32.5ms, p80:50ms, p95:50ms, iters:1088, tasks:28}, scan_detail: {total_process_keys: 1696, total_process_keys_size: 79129, total_keys: 1699, get_snapshot_time: 876.9µs, rocksdb: {key_skipped_count: 1696, block: {cache_hit_count: 15}}}, time_detail: {total_process_time: 2.43ms, total_suspend_time: 23.3µs, total_wait_time: 2.98ms, tikv_wall_time: 20.4ms} | not(isnull(test.t1.a)) | N/A | N/A |
| │ └─TableFullScan_29 | 1000000.00 | 1000000 | cop[tikv] | table:t1 | tikv_task:{proc max:50ms, min:0s, avg: 31.4ms, p80:50ms, p95:50ms, iters:1088, tasks:28} | keep order:false | N/A | N/A |
| └─StreamAgg_46(Probe) | 9965568.00 | 10000000 | root | | time:2.56s, loops:9767 | group by:test.t2.a, funcs:firstrow(test.t2.a)->test.t2.a | 8.39 KB | N/A |
| └─IndexReader_47 | 9965568.00 | 10000000 | root | | time:1.62s, loops:9774, cop_task: {num: 329, max: 61ms, min: 1.47ms, avg: 26.1ms, p95: 45.5ms, max_proc_keys: 50176, p95_proc_keys: 50176, tot_proc: 8s, tot_wait: 55.6ms, copr_cache_hit_ratio: 0.00, build_task_duration: 61.3µs, max_distsql_concurrency: 15}, rpc_info:{Cop:{num_rpc:329, total_time:8.59s}} | index:StreamAgg_38 | 52.8 MB | N/A |
| └─StreamAgg_38 | 9965568.00 | 10000000 | cop[tikv] | | tikv_task:{proc max:50ms, min:0s, avg: 23.8ms, p80:40ms, p95:40ms, iters:9774, tasks:329}, scan_detail: {total_process_keys: 10000000, total_process_keys_size: 460000000, total_keys: 10000329, get_snapshot_time: 26.3ms, rocksdb: {key_skipped_count: 10000000, block: {cache_hit_count: 16548}}}, time_detail: {total_process_time: 8s, total_suspend_time: 24.4ms, total_wait_time: 55.6ms, total_kv_read_wall_time: 7.28s, tikv_wall_time: 8.24s} | group by:test.t2.a, | N/A | N/A |
| └─IndexFullScan_45 | 10000000.00 | 10000000 | cop[tikv] | table:t2, index:idx_t2_a(a) | tikv_task:{proc max:50ms, min:0s, avg: 22.1ms, p80:40ms, p95:40ms, iters:9774, tasks:329} | keep order:true | N/A | N/A |
+------------------------------+-------------+----------+-----------+-----------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+----------------------------------------------------------+----------+---------+
8 rows in set (3.60 sec)
- Index Join 的优化效果
根据上述描述，执行计划默认选择走 Index Join 的情况下， 只能够使用外查询作为驱动表 ，因此在子查询比外查询结果集小时，执行速度可能会不及预期。t1 和 t2 仍然是 100 万 和 1000 万，通过执行计划可以发现，1000 万的 t2 (外表) 被作为驱动表进行 Build，语句执行 耗时达到 1分 31 秒 。
注：为了确保执行计划选择 Index Join，添加 /*+ no_hash_join(t2,t1@sel_2)*/ HINT。
mysql> explain analyze select /*+ no_hash_join(t2, t1@sel_2) */ * from t2 WHERE EXISTS (SELECT 1 FROM t1 WHERE t2.a = t1.a);
+------------------------------+-------------+----------+-----------+---------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+----------------------------------------------------------------------------------------------------------------+---------+------+
| id | estRows | actRows | task | access object | execution info | operator info | memory | disk |
+------------------------------+-------------+----------+-----------+---------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+----------------------------------------------------------------------------------------------------------------+---------+------+
| IndexJoin_14 | 8000000.00 | 1000000 | root | | time:1m31.2s, loops:978, RU:142580.352894, inner:{total:7m32.5s, concurrency:5, task:394, construct:11.5s, fetch:7m20.7s, build:303.2ms}, probe:1.46s | semi join, inner:IndexReader_13, outer key:test.t2.a, inner key:test.t1.a, equal cond:eq(test.t2.a, test.t1.a) | 31.2 MB | N/A |
| ├─TableReader_37(Build) | 10000000.00 | 10000000 | root | | time:50.7ms, loops:9779, cop_task: {num: 285, max: 72.9ms, min: 493.3µs, avg: 33.8ms, p95: 51.1ms, max_proc_keys: 50144, p95_proc_keys: 50144, tot_proc: 8.93s, tot_wait: 63.9ms, copr_cache_hit_ratio: 0.09, build_task_duration: 49.1µs, max_distsql_concurrency: 11}, rpc_info:{Cop:{num_rpc:285, total_time:9.61s}} | data:Selection_36 | 9.20 MB | N/A |
| │ └─Selection_36 | 10000000.00 | 10000000 | cop[tikv] | | tikv_task:{proc max:50ms, min:0s, avg: 30ms, p80:50ms, p95:50ms, iters:10898, tasks:285}, scan_detail: {total_process_keys: 9929120, total_process_keys_size: 466539933, total_keys: 9929380, get_snapshot_time: 39.6ms, rocksdb: {key_skipped_count: 9929120, block: {cache_hit_count: 19121}}}, time_detail: {total_process_time: 8.93s, total_suspend_time: 25.6ms, total_wait_time: 63.9ms, total_kv_read_wall_time: 8.13s, tikv_wall_time: 9.1s} | not(isnull(test.t2.a)) | N/A | N/A |
| │ └─TableFullScan_35 | 10000000.00 | 10000000 | cop[tikv] | table:t2 | tikv_task:{proc max:50ms, min:0s, avg: 28.8ms, p80:40ms, p95:50ms, iters:10898, tasks:285} | keep order:false | N/A | N/A |
| └─IndexReader_13(Probe) | 1003455.10 | 1000000 | root | | time:7m3.2s, loops:1436, cop_task: {num: 1202, max: 1.29s, min: 748.3µs, avg: 360.2ms, p95: 1.13s, max_proc_keys: 9184, p95_proc_keys: 5088, tot_proc: 6m33.2s, tot_wait: 209.3ms, copr_cache_hit_ratio: 0.03, build_task_duration: 580.2ms, max_distsql_concurrency: 2}, rpc_info:{Cop:{num_rpc:1202, total_time:7m12.9s}} | index:Selection_12 | 6.10 KB | N/A |
| └─Selection_12 | 1003455.10 | 1000000 | cop[tikv] | | tikv_task:{proc max:0s, min:0s, avg: 339.8ms, p80:901ms, p95:1.1s, iters:3594, tasks:1202}, scan_detail: {total_process_keys: 996040, total_process_keys_size: 45817840, total_keys: 10990023, get_snapshot_time: 57.4ms, rocksdb: {key_skipped_count: 996040, block: {cache_hit_count: 39977432}}}, time_detail: {total_process_time: 6m33.2s, total_suspend_time: 15.5s, total_wait_time: 209.3ms, total_kv_read_wall_time: 6m48.2s, tikv_wall_time: 7m6.6s} | not(isnull(test.t1.a)) | N/A | N/A |
| └─IndexRangeScan_11 | 1003455.10 | 1000000 | cop[tikv] | table:t1, index:idx_t1(a) | tikv_task:{proc max:0s, min:0s, avg: 339.8ms, p80:901ms, p95:1.1s, iters:3594, tasks:1202} | range: decided by [eq(test.t1.a, test.t2.a)], keep order:false | N/A | N/A |
+------------------------------+-------------+----------+-----------+---------------------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+----------------------------------------------------------------------------------------------------------------+---------+------+
7 rows in set (1 min 31.21 sec)
而当我们添加了 /*+ semi_join_rewrite() */ HINT 之后，执行计划发生了显著的变化。首先，驱动表变成 100 万的小表 t1，其次表的关联方式从 IndexJoin 变成了 IndexHashJoin。添加 HINT 后的执行耗时也是有了很大的改善，直接优化到 约 11 秒，提升了 9 倍。
mysql> explain analyze select /*+ no_hash_join(t2, t1@sel_2) */ * from t2 WHERE EXISTS (SELECT /*+ semi_join_rewrite() */ 1 FROM t1 WHERE t2.a = t1.a);
+----------------------------------+------------+---------+-----------+-----------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+-----------------------------------------------------------------------------------------------------------------+----------+------+
| id | estRows | actRows | task | access object | execution info | operator info | memory | disk |
+----------------------------------+------------+---------+-----------+-----------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+-----------------------------------------------------------------------------------------------------------------+----------+------+
| IndexHashJoin_20 | 1003390.88 | 1000000 | root | | time:10.8s, loops:978, RU:13383.173606, inner:{total:52.9s, concurrency:5, task:42, construct:935.8ms, fetch:51s, build:246.5ms, join:884.6ms} | inner join, inner:IndexLookUp_17, outer key:test.t1.a, inner key:test.t2.a, equal cond:eq(test.t1.a, test.t2.a) | 30.5 MB | N/A |
| ├─StreamAgg_72(Build) | 999936.00 | 1000000 | root | | time:98.2ms, loops:979 | group by:test.t1.a, funcs:firstrow(test.t1.a)->test.t1.a | 8.39 KB | N/A |
| │ └─IndexReader_73 | 999936.00 | 1000000 | root | | time:12.4ms, loops:978, cop_task: {num: 28, max: 3.2ms, min: 470µs, avg: 856µs, p95: 2.16ms, max_proc_keys: 1024, p95_proc_keys: 1024, tot_proc: 4.38ms, tot_wait: 3.47ms, copr_cache_hit_ratio: 0.86, build_task_duration: 22.3µs, max_distsql_concurrency: 1}, rpc_info:{Cop:{num_rpc:28, total_time:23.3ms}} | index:StreamAgg_68 | 6.36 MB | N/A |
| │ └─StreamAgg_68 | 999936.00 | 1000000 | cop[tikv] | | tikv_task:{proc max:50ms, min:0s, avg: 30.7ms, p80:40ms, p95:50ms, iters:977, tasks:28}, scan_detail: {total_process_keys: 4096, total_process_keys_size: 188416, total_keys: 4100, get_snapshot_time: 1.86ms, rocksdb: {key_skipped_count: 4096, block: {cache_hit_count: 22}}}, time_detail: {total_process_time: 4.38ms, total_suspend_time: 48.7µs, total_wait_time: 3.47ms, total_kv_read_wall_time: 10ms, tikv_wall_time: 12.5ms} | group by:test.t1.a, | N/A | N/A |
| │ └─IndexFullScan_36 | 1000000.00 | 1000000 | cop[tikv] | table:t1, index:idx_t1(a) | tikv_task:{proc max:50ms, min:0s, avg: 29.3ms, p80:40ms, p95:50ms, iters:977, tasks:28} | keep order:true | N/A | N/A |
| └─IndexLookUp_17(Probe) | 1003390.88 | 1000000 | root | | time:49.5s, loops:1019, index_task: {total_time: 49.3s, fetch_handle: 49.3s, build: 619.8µs, wait: 3.04ms}, table_task: {total_time: 915ms, num: 201, concurrency: 5}, next: {wait_index: 48.6s, wait_table_lookup_build: 99.4ms, wait_table_lookup_resp: 791.6ms} | | 158.1 KB | N/A |
| ├─Selection_16(Build) | 1003390.88 | 1000000 | cop[tikv] | | time:49.2s, loops:1142, cop_task: {num: 253, max: 666.6ms, min: 1.15ms, avg: 211.1ms, p95: 501.8ms, max_proc_keys: 9184, p95_proc_keys: 9184, tot_proc: 36.1s, tot_wait: 53.1ms, copr_cache_hit_ratio: 0.00, build_task_duration: 65.1ms, max_distsql_concurrency: 3}, rpc_info:{Cop:{num_rpc:253, total_time:53.4s}}, tikv_task:{proc max:650ms, min:0s, avg: 190.8ms, p80:300ms, p95:480ms, iters:2026, tasks:253}, scan_detail: {total_process_keys: 999840, total_process_keys_size: 45992640, total_keys: 1999853, get_snapshot_time: 29.8ms, rocksdb: {key_skipped_count: 999840, block: {cache_hit_count: 4787484}}}, time_detail: {total_process_time: 36.1s, total_suspend_time: 12.3s, total_wait_time: 53.1ms, total_kv_read_wall_time: 48.2s, tikv_wall_time: 52s} | not(isnull(test.t2.a)) | N/A | N/A |
| │ └─IndexRangeScan_14 | 1003390.88 | 1000000 | cop[tikv] | table:t2, index:idx_t2_a(a) | tikv_task:{proc max:650ms, min:0s, avg: 190.7ms, p80:300ms, p95:480ms, iters:2026, tasks:253} | range: decided by [eq(test.t2.a, test.t1.a)], keep order:false | N/A | N/A |
| └─TableRowIDScan_15(Probe) | 1003390.88 | 1000000 | cop[tikv] | table:t2 | time:775.3ms, loops:1258, cop_task: {num: 217, max: 26.9ms, min: 634.9µs, avg: 3.52ms, p95: 7.6ms, max_proc_keys: 8544, p95_proc_keys: 5024, tot_proc: 370.8ms, tot_wait: 32.1ms, copr_cache_hit_ratio: 0.33, build_task_duration: 10.7ms, max_distsql_concurrency: 2, max_extra_concurrency: 1}, rpc_info:{Cop:{num_rpc:217, total_time:756.1ms}}, tikv_task:{proc max:10ms, min:0s, avg: 3.92ms, p80:10ms, p95:10ms, iters:1992, tasks:217}, scan_detail: {total_process_keys: 381888, total_process_keys_size: 17893063, total_keys: 382135, get_snapshot_time: 12.6ms, rocksdb: {key_skipped_count: 381888, block: {cache_hit_count: 1681}}}, time_detail: {total_process_time: 370.8ms, total_suspend_time: 4.6ms, total_wait_time: 32.1ms, total_kv_read_wall_time: 370ms, tikv_wall_time: 475.2ms} | keep order:false | N/A | N/A |
+----------------------------------+------------+---------+-----------+-----------------------------+------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+-----------------------------------------------------------------------------------------------------------------+----------+------+
9 rows in set (10.83 sec)
关联子查询优化
关联子查询定义
关联子查询是指和外部查询有关联的子查询，即在子查询中使用了外部查询包含的列。在关联子查询中，对于外部查询返回的每一行数据，内部查询都要执行一次。
关联子查询具有以下几方面的特点：
- 信息流双向 ：关联子查询中的信息流是双向的。外部查询的每行数据传递一个值给子查询，子查询为每一行数据执行一次并返回记录，然后外部查询根据返回的记录做出决策。
- 灵活性 ：关联子查询可以使用关联列的灵活性，将SQL查询写成子查询的形式，这往往可以极大地简化SQL查询，并使SQL查询的语义更加方便理解。
- 执行挑战 ：为了计算关联结果的值（子查询的输出），关联子查询需要采用迭代（iterative）的执行方式。
以下是一个典型的关联子查询示例：
select * from t1 where t1.a < (select sum(t2.a) from t2 where t2.b = t1.b)；
TiDB 关联子查询优化-“子查询去关联”
关联子查询每次子查询执行时都是要和它的外部查询结果绑定的，因此，如果上述语句中 t1.a 有一千万个值，那这个子查询就要被重复执行一千万次，因为 t2.b=t1.b 这个条件会随着 t1.a 值的不同而发生变化。
如果能将关联依赖解除的话，这个子查询就只需要被执行一次。默认情况下，TiDB 会尝试进行子查询去关联，以达到更高的执行效率。以下是 TiDB 中针对这条语句生成的执行计划，可以看出 TiDB 的优化器将语句的执行计划生成了一个 Hash Join 关联方式。模拟2个表的数据量为 1千万行进行测试时，语句的实际执行耗时 约为 14 秒 。
mysql> explain select * from t1 where t1.a < (select sum(t2.a) from t2 where t2.b = t1.b);
+--------------------------------+-----------+-----------+---------------+--------------------------------------------------------------------------------------------------------------+
| id | estRows | task | access object | operator info |
+--------------------------------+-----------+-----------+---------------+--------------------------------------------------------------------------------------------------------------+
| HashJoin_12 | 452412.14 | root | | inner join, equal:[eq(test.t1.b, test.t2.b)], other cond:lt(cast(test.t1.a, decimal(10,0) BINARY), Column#7) |
| ├─TableReader_15(Build) | 452412.14 | root | | data:Selection_14 |
| │ └─Selection_14 | 452412.14 | cop[tikv] | | not(isnull(test.t1.b)) |
| │ └─TableFullScan_13 | 452865.00 | cop[tikv] | table:t1 | keep order:false, stats:pseudo |
| └─HashAgg_21(Probe) | 472411.12 | root | | group by:test.t2.b, funcs:sum(Column#8)->Column#7, funcs:firstrow(test.t2.b)->test.t2.b |
| └─TableReader_22 | 472411.12 | root | | data:HashAgg_16 |
| └─HashAgg_16 | 472411.12 | cop[tikv] | | group by:test.t2.b, funcs:sum(test.t2.a)->Column#8 |
| └─Selection_20 | 590513.90 | cop[tikv] | | not(isnull(test.t2.b)) |
| └─TableFullScan_19 | 591105.00 | cop[tikv] | table:t2 | keep order:false, stats:pseudo |
+--------------------------------+-----------+-----------+---------------+--------------------------------------------------------------------------------------------------------------+
9 rows in set (0.00 sec)
假如 TiDB 没有做子查询去关联的优化，实际执行情况又是怎么样呢？我们可以用 HINT /*+ NO_DECORRELATE() */ 来关闭子查询去关联，模拟未优化前的情况。此时得到如下的执行计划，其中 operator info 为 range: decided by [eq(test.t2.b, test.t1.b)] 的 IndexRangeScan_42 算子表示TiDB 使用关联条件进行索引范围查询。同样的语句执行耗时 超过 1 小时 。
mysql> explain select * from t1 where t1.a < (select /*+ NO_DECORRELATE() */ sum(t2.a) from t2 where t2.b = t1.b);
+------------------------------------------+-----------+-----------+---------------------------+--------------------------------------------------------------------------------------+
| id | estRows | task | access object | operator info |
+------------------------------------------+-----------+-----------+---------------------------+--------------------------------------------------------------------------------------+
| Projection_10 | 452865.00 | root | | test.t1.a, test.t1.b |
| └─Apply_12 | 452865.00 | root | | CARTESIAN inner join, other cond:lt(cast(test.t1.a, decimal(10,0) BINARY), Column#7) |
| ├─TableReader_14(Build) | 452865.00 | root | | data:TableFullScan_13 |
| │ └─TableFullScan_13 | 452865.00 | cop[tikv] | table:t1 | keep order:false, stats:pseudo |
| └─MaxOneRow_15(Probe) | 452865.00 | root | | |
| └─StreamAgg_20 | 452865.00 | root | | funcs:sum(Column#14)->Column#7 |
| └─Projection_45 | 857574.13 | root | | cast(test.t2.a, decimal(10,0) BINARY)->Column#14 |
| └─IndexLookUp_44 | 857574.13 | root | | |
| ├─IndexRangeScan_42(Build) | 857574.13 | cop[tikv] | table:t2, index:idx_t2(b) | range: decided by [eq(test.t2.b, test.t1.b)], keep order:false |
| └─TableRowIDScan_43(Probe) | 857574.13 | cop[tikv] | table:t2 | keep order:false |
+------------------------------------------+-----------+-----------+---------------------------+--------------------------------------------------------------------------------------+
10 rows in set (0.01 sec)
何时需要关闭 “子查询去关联”
那么，是不是所有的关联子查询语句在关联依赖解除优化后性能都能有较大的提升呢？并不是这样。仍然用上面的示例来说，在外部的值比较少的情况下，不解除关联依赖反而可能对执行性能更有帮助。例如，如果 t3 表只有 100 条记录时，以下语句执行耗时 仅 0.28 秒 。
mysql> explain analyze select * from t3 where t3.a = (select /*+ NO_DECORRELATE() */ sum(t2.a) from t2 where t2.b = t3.b);
+------------------------------------------+---------+---------+-----------+---------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+-----------------------------------------------------------------------+-----------+------+
| id | estRows | actRows | task | access object | execution info | operator info | memory | disk |
+------------------------------------------+---------+---------+-----------+---------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+-----------------------------------------------------------------------+-----------+------+
| Projection_11 | 100.00 | 0 | root | | time:277.9ms, loops:1, RU:117.173904, Concurrency:OFF | test.t3.a, test.t3.b | 3.55 KB | N/A |
| └─Apply_13 | 100.00 | 0 | root | | time:277.9ms, loops:1, concurrency:OFF, cache:ON, cacheHitRatio:0.000% | inner join, equal:[eq(Column#8, Column#7)] | 900 Bytes | N/A |
| ├─Projection_14(Build) | 100.00 | 100 | root | | time:1.61ms, loops:2, Concurrency:OFF | test.t3.a, test.t3.b, cast(test.t3.a, decimal(10,0) BINARY)->Column#8 | 1.82 KB | N/A |
| │ └─TableReader_16 | 100.00 | 100 | root | | time:1.56ms, loops:2, cop_task: {num: 1, max: 1.36ms, proc_keys: 100, tot_proc: 333.9µs, tot_wait: 146.6µs, copr_cache_hit_ratio: 0.00, build_task_duration: 18.3µs, max_distsql_concurrency: 1}, rpc_info:{Cop:{num_rpc:1, total_time:1.32ms}} | data:TableFullScan_15 | 1.86 KB | N/A |
| │ └─TableFullScan_15 | 100.00 | 100 | cop[tikv] | table:t3 | tikv_task:{time:0s, loops:3}, scan_detail: {total_process_keys: 100, total_process_keys_size: 4400, total_keys: 101, get_snapshot_time: 46.4µs, rocksdb: {delete_skipped_count: 100, key_skipped_count: 200, block: {cache_hit_count: 4}}}, time_detail: {total_process_time: 333.9µs, total_wait_time: 146.6µs, tikv_wall_time: 730.8µs} | keep order:false, stats:pseudo | N/A | N/A |
| └─MaxOneRow_17(Probe) | 100.00 | 100 | root | | time:265.2ms, loops:200 | | N/A | N/A |
| └─StreamAgg_22 | 100.00 | 100 | root | | time:264.7ms, loops:200 | funcs:sum(Column#16)->Column#7 | 8.75 KB | N/A |
| └─Projection_47 | 189.37 | 178 | root | | time:264ms, loops:176, Concurrency:OFF | cast(test.t2.a, decimal(10,0) BINARY)->Column#16 | 8.61 KB | N/A |
| └─IndexLookUp_46 | 189.37 | 178 | root | | time:263.1ms, loops:176, index_task: {total_time: 78.3ms, fetch_handle: 77.5ms, build: 124.7µs, wait: 629µs}, table_task: {total_time: 144.4ms, num: 76, concurrency: 5}, next: {wait_index: 90ms, wait_table_lookup_build: 7.42ms, wait_table_lookup_resp: 136.1ms} | | 26.4 KB | N/A |
| ├─IndexRangeScan_44(Build) | 189.37 | 178 | cop[tikv] | table:t2, index:idx_t2(b) | time:100ms, loops:252, cop_task: {num: 100, max: 1.38ms, min: 596.8µs, avg: 880.7µs, p95: 1.2ms, max_proc_keys: 8, p95_proc_keys: 5, tot_proc: 13.5ms, tot_wait: 9.62ms, copr_cache_hit_ratio: 0.00, build_task_duration: 2.07ms, max_distsql_concurrency: 1}, rpc_info:{Cop:{num_rpc:100, total_time:85.1ms}}, tikv_task:{proc max:10ms, min:0s, avg: 300µs, p80:0s, p95:0s, iters:100, tasks:100}, scan_detail: {total_process_keys: 178, total_process_keys_size: 8188, total_keys: 278, get_snapshot_time: 3.17ms, rocksdb: {key_skipped_count: 178, block: {cache_hit_count: 465}}}, time_detail: {total_process_time: 13.5ms, total_wait_time: 9.62ms, total_kv_read_wall_time: 30ms, tikv_wall_time: 41.8ms} | range: decided by [eq(test.t2.b, test.t3.b)], keep order:false | N/A | N/A |
| └─TableRowIDScan_45(Probe) | 189.37 | 178 | cop[tikv] | table:t2 | time:133.3ms, loops:152, cop_task: {num: 159, max: 9.38ms, min: 0s, avg: 753.7µs, p95: 1.13ms, max_proc_keys: 2, p95_proc_keys: 2, tot_proc: 18.6ms, tot_wait: 24.3ms, copr_cache_hit_ratio: 0.00, build_task_duration: 3.5ms, max_distsql_concurrency: 1, max_extra_concurrency: 1, store_batch_num: 34}, rpc_info:{Cop:{num_rpc:125, total_time:117ms}}, tikv_task:{proc max:10ms, min:0s, avg: 62.9µs, p80:0s, p95:0s, iters:159, tasks:159}, scan_detail: {total_process_keys: 178, total_process_keys_size: 8366, total_keys: 178, get_snapshot_time: 5.14ms, rocksdb: {block: {cache_hit_count: 798}}}, time_detail: {total_process_time: 18.6ms, total_wait_time: 24.3ms, total_kv_read_wall_time: 10ms, tikv_wall_time: 63.3ms} | keep order:false | N/A | N/A |
+------------------------------------------+---------+---------+-----------+---------------------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+-----------------------------------------------------------------------+-----------+------+
11 rows in set (0.28 sec)
而如果走 TiDB 默认的解除关联依赖时，耗时 约 5 秒 ，且存在 OOM 的风险。
mysql> explain analyze select * from t3 where t3.a = (select sum(t2.a) from t2 where t2.b = t3.b);
ERROR 8175 (HY000): Your query has been cancelled due to exceeding the allowed memory limit for a single SQL query. Please try narrowing your query scope or increase the tidb_mem_quota_query limit and try again.[conn=480325804]
mysql> set tidb_mem_quota_query=10737418240;
Query OK, 0 rows affected (0.00 sec)mysql> explain analyze select * from t3 where t3.a = (select sum(t2.a) from t2 where t2.b = t3.b);
+--------------------------------+-------------+----------+-----------+---------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+------------------------------------------------------------------------------------------+---------+---------+
| id | estRows | actRows | task | access object | execution info | operator info | memory | disk |
+--------------------------------+-------------+----------+-----------+---------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+------------------------------------------------------------------------------------------+---------+---------+
| HashJoin_13 | 99.90 | 0 | root | | time:5.11s, loops:1, RU:11889.809066, build_hash_table:{total:1.48ms, fetch:1.36ms, build:122.2µs}, probe:{concurrency:5, total:25.5s, max:5.11s, probe:1.44s, fetch and wait:24.1s} | inner join, equal:[eq(test.t3.b, test.t2.b) eq(Column#8, Column#7)] | 46.0 KB | 0 Bytes |
| ├─Projection_14(Build) | 99.90 | 100 | root | | time:1.26ms, loops:2, Concurrency:OFF | test.t3.a, test.t3.b, cast(test.t3.a, decimal(10,0) BINARY)->Column#8 | 2.52 KB | N/A |
| │ └─TableReader_17 | 99.90 | 100 | root | | time:1.22ms, loops:2, cop_task: {num: 1, max: 1.38ms, proc_keys: 100, tot_proc: 353µs, tot_wait: 129.8µs, copr_cache_hit_ratio: 0.00, build_task_duration: 24.3µs, max_distsql_concurrency: 1}, rpc_info:{Cop:{num_rpc:1, total_time:1.33ms}} | data:Selection_16 | 1.87 KB | N/A |
| │ └─Selection_16 | 99.90 | 100 | cop[tikv] | | tikv_task:{time:0s, loops:3}, scan_detail: {total_process_keys: 100, total_process_keys_size: 4400, total_keys: 101, get_snapshot_time: 49.9µs, rocksdb: {delete_skipped_count: 100, key_skipped_count: 200, block: {cache_hit_count: 4}}}, time_detail: {total_process_time: 353µs, total_wait_time: 129.8µs, tikv_wall_time: 716.5µs} | not(isnull(test.t3.b)) | N/A | N/A |
| │ └─TableFullScan_15 | 100.00 | 100 | cop[tikv] | table:t3 | tikv_task:{time:0s, loops:3} | keep order:false, stats:pseudo | N/A | N/A |
| └─HashAgg_28(Probe) | 5280768.00 | 5343857 | root | | time:5.09s, loops:5222, partial_worker:{wall_time:2.494534938s, concurrency:5, task_num:285, tot_wait:215.654904ms, tot_exec:12.25579073s, tot_time:12.47193031s, max:2.494394166s, p95:2.494394166s}, final_worker:{wall_time:5.107395873s, concurrency:5, task_num:25, tot_wait:70.452µs, tot_exec:8.250691502s, tot_time:24.968242542s, max:5.107349993s, p95:5.107349993s} | group by:test.t2.b, funcs:sum(Column#11)->Column#7, funcs:firstrow(test.t2.b)->test.t2.b | 1.95 GB | 0 Bytes |
| └─TableReader_29 | 5280768.00 | 9972967 | root | | time:225.2ms, loops:286, cop_task: {num: 285, max: 157.9ms, min: 470.7µs, avg: 55.1ms, p95: 83.7ms, max_proc_keys: 50176, p95_proc_keys: 50176, tot_proc: 13.8s, tot_wait: 51ms, copr_cache_hit_ratio: 0.04, build_task_duration: 41.8µs, max_distsql_concurrency: 11}, rpc_info:{Cop:{num_rpc:285, total_time:15.7s}} | data:HashAgg_21 | 25.2 MB | N/A |
| └─HashAgg_21 | 5280768.00 | 9972967 | cop[tikv] | | tikv_task:{proc max:90ms, min:0s, avg: 46.7ms, p80:70ms, p95:70ms, iters:9769, tasks:285}, scan_detail: {total_process_keys: 9982592, total_process_keys_size: 469050751, total_keys: 9982867, get_snapshot_time: 19.6ms, rocksdb: {delete_skipped_count: 874, key_skipped_count: 9983466, block: {cache_hit_count: 18229}}}, time_detail: {total_process_time: 13.8s, total_suspend_time: 37.8ms, total_wait_time: 51ms, total_kv_read_wall_time: 9.13s, tikv_wall_time: 14.3s} | group by:test.t2.b, funcs:sum(test.t2.a)->Column#11 | N/A | N/A |
| └─Selection_27 | 10000000.00 | 10000000 | cop[tikv] | | tikv_task:{proc max:70ms, min:0s, avg: 33.2ms, p80:50ms, p95:60ms, iters:9769, tasks:285} | not(isnull(test.t2.b)) | N/A | N/A |
| └─TableFullScan_26 | 10000000.00 | 10000000 | cop[tikv] | table:t2 | tikv_task:{proc max:70ms, min:0s, avg: 32ms, p80:50ms, p95:60ms, iters:9769, tasks:285} | keep order:false | N/A | N/A |
+--------------------------------+-------------+----------+-----------+---------------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+------------------------------------------------------------------------------------------+---------+---------+
10 rows in set (5.12 sec)
全局关闭 “子查询去关联”
/*+ NO_DECORRELATE() */ HINT 是一种针对具体 SQL 级别来关闭子查询去关联的方式，TiDB 也提供另外一种可以全局关闭的方式，它是直接将子查询去关联的规则加入到黑名单中。
关于优化规则黑名单的使用，可以参考 TiDB 官网文档 https://docs.pingcap.com/zh/tidb/v7.5/blocklist-control-plan。如果希望全局关闭子查询去关联的功能，需要执行以下命令，此后不需要使用 /*+ NO_DECORRELATE() */ 这个 HINT 我们也同样可以获得具有子查询关联的执行计划。
insert into mysql.opt_rule_blacklist values("decorrelate");
admin reload opt_rule_blacklist;
优化总结
TiDB 优化器针对 semi join 半连接默认会转换为 HashJoin 或 IndexJoin 以获得更好的性能，然而在某些特殊的场景下默认选择的 HashJoin 或 IndexJoin 可能仍然达不到最佳的性能。TiDB 提供的 SEMI_JOIN_REWRITE() 可以扩大选择范围，让优化器选择更好的执行方式，从而在性能上达到几倍的性能。
TiDB 默认针对关联子查询有自动解除关联依赖的优化操作，当涉及查询条件的数据量较大时，将重复执行无数次的子查询转化为只需执行一次，从而大大缩减了执行耗时。然而，如果满足查询条件的数据量很小的话，这种优化方式可能会造成相反的效果，此时我们则需要借助 NO_DECORRELATE() 来关闭 “子查询去关联” 的优化，也可以通过全局添加规则黑名单的方式实现全局控制。
