## 问题描述 bug 触发条件如下: 1. 优化器先选择了 where 条件中字段的索引,该索引过滤性较好; 2. SQL 中必须有 order by limit 从而引导优化器尝试使用 order by 字段上的索引进行优化,最终因代价问题没有成功。 ### 复现case 表结构 ~~~ create table t1( id int auto_increment primary key, a int, b int, c int, key iabc (a, b, c), key ic (c) ) engine = innodb; ~~~ 构造数据 ~~~ insert into t1 select null,null,null,null; insert into t1 select null,null,null,null from t1; insert into t1 select null,null,null,null from t1; insert into t1 select null,null,null,null from t1; insert into t1 select null,null,null,null from t1; insert into t1 select null,null,null,null from t1; update t1 set a = id / 2, b = id / 4, c = 6 - id / 8; ~~~ 触发SQL ~~~ mysql> explain select id from t1 where a<3 and b in (1, 13) and c>=3 order by c limit 2\G *************************** 1\. row *************************** id: 1 select_type: SIMPLE table: t1 type: index possible_keys: iabc,ic key: iabc key_len: 15 ref: NULL rows: 32 Extra: Using where; Using index; Using filesort ~~~ 使用 force index 可以选择过滤性好的索引 ~~~ mysql> explain select id from t1 force index(iabc) where a<3 and b in (1, 13) and c>=3 order by c limit 2\G *************************** 1\. row *************************** id: 1 select_type: SIMPLE table: t1 type: range possible_keys: iabc key: iabc key_len: 5 ref: NULL rows: 3 Extra: Using where; Using index; Using filesort ~~~ ## 问题分析 optimizer_trace 可以帮助分析这个问题。 SELECT * FROM INFORMATION_SCHEMA.OPTIMIZER_TRACE\G ~~~ "range_scan_alternatives": [ { "index": "iabc", "ranges": [ "NULL < a < 3" ], "index_dives_for_eq_ranges": true, "rowid_ordered": false, "using_mrr": false, "index_only": true, "rows": 3, "cost": 1.6146, "chosen": true }, { "index": "ic", "ranges": [ "3 <= c" ], "index_dives_for_eq_ranges": true, "rowid_ordered": false, "using_mrr": false, "index_only": false, "rows": 17, "cost": 21.41, "chosen": false, "cause": "cost" } ], ~~~ range_scan_alternatives 计算 range_scan,各个索引的开销,从上面的结果可以看出,联合索引 iabc 开销较小,应该选择 iabc。 ~~~ "considered_execution_plans": [ { "plan_prefix": [ ], "table": "`t1`", "best_access_path": { "considered_access_paths": [ { "access_type": "range", "rows": 3, "cost": 2.2146, "chosen": true } ] }, "cost_for_plan": 2.2146, "rows_for_plan": 3, "chosen": true } ] ~~~ considered_execution_plans 表索引选择过程,access_type 是 range,rows_for_plan=3,到这里为止,执行计划还是符合预期的。 ~~~ { "clause_processing": { "clause": "ORDER BY", "original_clause": "`t1`.`c`", "items": [ { "item": "`t1`.`c`" } ], "resulting_clause_is_simple": true, "resulting_clause": "`t1`.`c`" } }, { "refine_plan": [ { "table": "`t1`", "access_type": "index_scan" } ] }, { "reconsidering_access_paths_for_index_ordering": { "clause": "ORDER BY", "index_order_summary": { "table": "`t1`", "index_provides_order": false, "order_direction": "undefined", "index": "unknown", "plan_changed": false } } } ~~~ clause_processing 用于简化 order by,经过 clause_processing access_type 变成 index_scan(全索引扫描,过滤性较range差),此时出现了和预期不符的结果。 因此可以推测优化器试图优化 order by 时出现了错误: * 第一阶段,优化器选择了索引 iabc,采用 range 访问; * 第二阶段,优化器试图进一步优化执行计划,使用 order by 的列访问,并清空了第一阶段的结果; * 第三阶段,优化器发现使用 order by 的列访问,代价比第一阶段的结果更大,但是第一阶段结果已经被清空了,无法还原,于是选择了代价较大的访问方式(index_scan),触发了bug。 ## 问题解决 1. 我们在索引优化函数`SQL_SELECT::test_quick_select` 最开始的时候保存访问计划变量(quick); 2. 在索引没变的时候,还原这个变量; 3. 在索引发生改变的时候,删除这个变量。 在不修改 mysql 源码的情况下,可以通过 force index 强制指定索引规避这个bug。 `SQL_SELECT::test_quick_select` 调用栈如下 ~~~ #0 SQL_SELECT::test_quick_select #1 make_join_select #2 JOIN::optimize #3 mysql_execute_select #4 mysql_select #5 mysql_explain_unit #6 explain_query_expression #7 execute_sqlcom_select #8 mysql_execute_command #9 mysql_parse #10 dispatch_command #11 do_command ~~~