在开始演示之前,我们先介绍下两个概念。
概念一,数据的可选择性基数,也就是常说的cardinality值。
查询优化器在生成各种执行计划之前,得先从统计信息中取得相关数据,这样才能估算每步操作所涉及到的记录数,而这个相关数据就是cardinality。简单来说,就是每个值在每个字段中的唯一值分布状态。
比如表t1有100行记录,其中一列为f1。f1中唯一值的个数可以是100个,也可以是1个,当然也可以是1到100之间的任何一个数字。这里唯一值越的多少,就是这个列的可选择基数。
那看到这里我们就明白了,为什么要在基数高的字段上建立索引,而基数低的的字段建立索引反而没有全表扫描来的快。当然这个只是一方面,至于更深入的探讨就不在我这篇探讨的范围了。
概念二,关于HINT的使用。
这里我来说下HINT是什么,在什么时候用。
HINT简单来说就是在某些特定的场景下人工协助MySQL优化器的工作,使她生成最优的执行计划。一般来说,优化器的执行计划都是最优化的,不过在某些特定场景下,执行计划可能不是最优化。
比如:表t1经过大量的频繁更新操作,(UPDATE,DELETE,INSERT),cardinality已经很不准确了,这时候刚好执行了一条SQL,那么有可能这条SQL的执行计划就不是最优的。为什么说有可能呢?
来看下具体演示
譬如,以下两条SQL,
A:
select * from t1 where f1 = 20;B:
select * from t1 where f1 = 30;如果f1的值刚好频繁更新的值为30,并且没有达到MySQL自动更新cardinality值的临界值或者说用户设置了手动更新又或者用户减少了sample page等等,那么对这两条语句来说,可能不准确的就是B了。
这里顺带说下,MySQL提供了自动更新和手动更新表cardinality值的方法,因篇幅有限,需要的可以查阅手册。
那回到正题上,MySQL 8.0 带来了几个HINT,我今天就举个index_merge的例子。
示例表结构:
mysql> desc t1;+------------+--------------+------+-----+---------+----------------+| Field | Type | Null | Key | Default | Extra |+------------+--------------+------+-----+---------+----------------+| id | int(11) | NO | PRI | NULL | auto_increment || rank1 | int(11) | YES | MUL | NULL | || rank2 | int(11) | YES | MUL | NULL | || log_time | datetime | YES | MUL | NULL | || prefix_uid | varchar(100) | YES | | NULL | || desc1 | text | YES | | NULL | || rank3 | int(11) | YES | MUL | NULL | |+------------+--------------+------+-----+---------+----------------+7 rows in set (0.00 sec)表记录数:
这里我们两条经典的SQL:
SQL C:
select * from t1 where rank1 = 1 or rank2 = 2 or rank3 = 2;SQL D:
select * from t1 where rank1 =100 and rank2 =100 and rank3 =100;表t1实际上在rank1,rank2,rank3三列上分别有一个二级索引。
那我们来看SQL C的查询计划。
显然,没有用到任何索引,扫描的行数为32034,cost为3243.65。
mysql> explain format=json select * from t1 where rank1 =1 or rank2 = 2 or rank3 = 2\G*************************** 1. row ***************************EXPLAIN: { "query_block": { "select_id": 1, "cost_info": { "query_cost": "3243.65" }, "table": { "table_name": "t1", "access_type": "ALL", "possible_keys": [ "idx_rank1", "idx_rank2", "idx_rank3" ], "rows_examined_per_scan": 32034, "rows_produced_per_join": 115, "filtered": "0.36", "cost_info": { "read_cost": "3232.07", "eval_cost": "11.58", "prefix_cost": "3243.65", "data_read_per_join": "49K" }, "used_columns": [ "id", "rank1", "rank2", "log_time", "prefix_uid", "desc1", "rank3" ], "attached_condition": "((`ytt`.`t1`.`rank1` = 1) or (`ytt`.`t1`.`rank2` = 2) or (`ytt`.`t1`.`rank3` = 2))" } }}1 row in set, 1 warning (0.00 sec)我们加上hint给相同的查询,再次看看查询计划。
这个时候用到了index_merge,union了三个列。扫描的行数为1103,cost为441.09,明显比之前的快了好几倍。
mysql> explain format=json select /*+ index_merge(t1) */ * from t1 where rank1 =1 or rank2 = 2 or rank3 = 2\G*************************** 1. row ***************************EXPLAIN: { "query_block": { "select_id": 1, "cost_info": { "query_cost": "441.09" }, "table": { "table_name": "t1", "access_type": "index_merge", "possible_keys": [ "idx_rank1", "idx_rank2", "idx_rank3" ], "key": "union(idx_rank1,idx_rank2,idx_rank3)", "key_length": "5,5,5", "rows_examined_per_scan": 1103, "rows_produced_per_join": 1103, "filtered": "100.00", "cost_info": { "read_cost": "330.79", "eval_cost": "110.30", "prefix_cost": "441.09", "data_read_per_join": "473K" }, "used_columns": [ "id", "rank1", "rank2", "log_time", "prefix_uid", "desc1", "rank3" ], "attached_condition": "((`ytt`.`t1`.`rank1` = 1) or (`ytt`.`t1`.`rank2` = 2) or (`ytt`.`t1`.`rank3` = 2))" } }}1 row in set, 1 warning (0.00 sec)我们再看下SQL D的计划:
不加HINT,
mysql> explain format=json select * from t1 where rank1 =100 and rank2 =100 and rank3 =100\G*************************** 1. row ***************************EXPLAIN: { "query_block": { "select_id": 1, "cost_info": { "query_cost": "534.34" }, "table": { "table_name": "t1", "access_type": "ref", "possible_keys": [ "idx_rank1", "idx_rank2", "idx_rank3" ], "key": "idx_rank1", "used_key_parts": [ "rank1" ], "key_length": "5", "ref": [ "const" ], "rows_examined_per_scan": 555, "rows_produced_per_join": 0, "filtered": "0.07", "cost_info": { "read_cost": "478.84", "eval_cost": "0.04", "prefix_cost": "534.34", "data_read_per_join": "176" }, "used_columns": [ "id", "rank1", "rank2", "log_time", "prefix_uid", "desc1", "rank3" ], "attached_condition": "((`ytt`.`t1`.`rank3` = 100) and (`ytt`.`t1`.`rank2` = 100))" } }}1 row in set, 1 warning (0.00 sec)加了HINT,
mysql> explain format=json select /*+ index_merge(t1)*/ * from t1 where rank1 =100 and rank2 =100 and rank3 =100\G*************************** 1. row ***************************EXPLAIN: { "query_block": { "select_id": 1, "cost_info": { "query_cost": "5.23" }, "table": { "table_name": "t1", "access_type": "index_merge", "possible_keys": [ "idx_rank1", "idx_rank2", "idx_rank3" ], "key": "intersect(idx_rank1,idx_rank2,idx_rank3)", "key_length": "5,5,5", "rows_examined_per_scan": 1, "rows_produced_per_join": 1, "filtered": "100.00", "cost_info": { "read_cost": "5.13", "eval_cost": "0.10", "prefix_cost": "5.23", "data_read_per_join": "440" }, "used_columns": [ "id", "rank1", "rank2", "log_time", "prefix_uid", "desc1", "rank3" ], "attached_condition": "((`ytt`.`t1`.`rank3` = 100) and (`ytt`.`t1`.`rank2` = 100) and (`ytt`.`t1`.`rank1` = 100))" } }}1 row in set, 1 warning (0.00 sec)对比下以上两个,加了HINT的比不加HINT的cost小了100倍。
总结下,就是说表的cardinality值影响这张的查询计划,如果这个值没有正常更新的话,就需要手工加HINT了。相信MySQL未来的版本会带来更多的HINT。