MySQL 5.6 查询优化器新特性的“BUG”

最近碰到一个慢SQL问题,解决过程有点小曲折,和大家分享下。 SQL本身不复杂,表结构、索引也比较简单,不过个别字段存在于多个索引中。

CREATE TABLE `pre_forum_post` (
  `pid` int(10) unsigned NOT NULL,
  `fid` mediumint(8) unsigned NOT NULL DEFAULT '0',
  `tid` mediumint(8) unsigned NOT NULL DEFAULT '0',
  `first` tinyint(1) NOT NULL DEFAULT '0',
  `author` varchar(40) NOT NULL DEFAULT '',
  `authorid` int(10) unsigned NOT NULL DEFAULT '0',
  `subject` varchar(80) NOT NULL DEFAULT '',
  `dateline` int(10) unsigned NOT NULL DEFAULT '0',
  `message` mediumtext NOT NULL,
  `useip` varchar(15) NOT NULL DEFAULT '',
  `invisible` tinyint(1) NOT NULL DEFAULT '0',
  `anonymous` tinyint(1) NOT NULL DEFAULT '0',
  `usesig` tinyint(1) NOT NULL DEFAULT '0',
  `htmlon` tinyint(1) NOT NULL DEFAULT '0',
  `bbcodeoff` tinyint(1) NOT NULL DEFAULT '0',
  `smileyoff` tinyint(1) NOT NULL DEFAULT '0',
  `parseurloff` tinyint(1) NOT NULL DEFAULT '0',
  `attachment` tinyint(1) NOT NULL DEFAULT '0',
  `rate` smallint(6) NOT NULL DEFAULT '0',
  `ratetimes` tinyint(3) unsigned NOT NULL DEFAULT '0',
  `status` int(10) NOT NULL DEFAULT '0',
  `tags` varchar(255) NOT NULL DEFAULT '0',
  `comment` tinyint(1) NOT NULL DEFAULT '0',
  `replycredit` int(10) NOT NULL DEFAULT '0',
  `position` int(8) unsigned NOT NULL AUTO_INCREMENT,
  PRIMARY KEY (`tid`,`position`),
  UNIQUE KEY `pid` (`pid`),
  KEY `fid` (`fid`),
  KEY `displayorder` (`tid`,`invisible`,`dateline`),
  KEY `first` (`tid`,`first`),
  KEY `new_auth` (`authorid`,`invisible`,`tid`),
  KEY `idx_dt` (`dateline`)
) ENGINE=MyISAM DEFAULT CHARSET=utf8;


"root@localhost Fri Aug  1 11:59:56 2014 11:59:56 [test]>show table status like 'pre_forum_post'\G
*************************** 1. row ***************************
           Name: pre_forum_post
         Engine: MyISAM
        Version: 10
     Row_format: Dynamic
           Rows: 23483977
 Avg_row_length: 203
    Data_length: 4782024708
Max_data_length: 281474976710655
   Index_length: 2466093056
      Data_free: 0
 Auto_increment: 1
    Create_time: 2014-08-01 11:00:56
    Update_time: 2014-08-01 11:08:49
     Check_time: 2014-08-01 11:12:23
      Collation: utf8_general_ci
       Checksum: NULL
 Create_options: 
        Comment: 


mysql> show index from pre_forum_post;
+----------------+------------+--------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+
| Table          | Non_unique | Key_name     | Seq_in_index | Column_name | Collation | Cardinality | Sub_part | Packed | Null | Index_type | Comment | Index_comment |
+----------------+------------+--------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+
| pre_forum_post |          0 | PRIMARY      |            1 | tid         | A         |      838713 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          0 | PRIMARY      |            2 | position    | A         |    23483977 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          0 | pid          |            1 | pid         | A         |    23483977 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          1 | fid          |            1 | fid         | A         |        1470 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          1 | displayorder |            1 | tid         | A         |      838713 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          1 | displayorder |            2 | invisible   | A         |      869776 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          1 | displayorder |            3 | dateline    | A         |    23483977 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          1 | first        |            1 | tid         | A         |      838713 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          1 | first        |            2 | first       | A         |     1174198 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          1 | new_auth     |            1 | authorid    | A         |     1806459 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          1 | new_auth     |            2 | invisible   | A         |     1956998 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          1 | new_auth     |            3 | tid         | A         |    11741988 |     NULL | NULL   |      | BTREE      |         |               |
| pre_forum_post |          1 | idx_dt       |            1 | dateline    | A         |    23483977 |     NULL | NULL   |      | BTREE      |         |               |
+----------------+------------+--------------+--------------+-------------+-----------+-------------+----------+--------+------+------------+---------+---------------+

我们来看下这个SQL的执行计划:

mysql> explain select * from pre_forum_post where tid=7932612 and `invisible` in('0','-2') order by dateline  limit 15\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: pre_forum_post
         type: index
possible_keys: PRIMARY,displayorder,first
          key: idx_dt
      key_len: 4
          ref: NULL
         rows: 14042
        Extra: Using where

可以看到执行计划比较奇怪,从几个可选的索引中,最终选择了 idx_dt,结果悲剧了,这个SQL执行耗时很长:

mysql> select * from pre_forum_post where tid=7932612 and `invisible` in('0','-2') order by dateline  limit 15;
15 rows in set (26.78 sec)

看下MySQL的会话状态值:Handler_read_next

| Handler_read_next          | 17274153 |

1700多万数据中选取15条记录,结果可想而知,非常慢。 我们强制指定比较靠谱的索引再看下:

mysql> explain select * from pre_forum_post force index(displayorder) where tid=7932612 and `invisible` in('0','-2') order by dateline  limit 15\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: pre_forum_post
         type: range
possible_keys: displayorder
          key: displayorder
      key_len: 4
          ref: NULL
         rows: 46131
        Extra: Using index condition; Using filesort

看下实际执行的耗时:

mysql> select * from pre_forum_post force index(displayorder) where tid=7932612 and `invisible` in('0','-2') order by dateline  limit 15;
15 rows in set (0.08 sec)

尼玛,怎么可以这么快,查询优化器未免太坑爹了吧。 再看下MySQL的会话状态值:Handler_read_next

| Handler_read_next          | 31188 |

和不强制索引的情况相比,差了553倍! 所幸,5.6以上除了EXPLAIN外,还支持OPTIMIZER_TRACE,我们来观察下两种执行计划的区别,发现不强制指定索引时的执行计划有诈,会在最后判断到 ORDER BY 子句时,修改执行计划:

          {\
            "reconsidering_access_paths_for_index_ordering": {\
              "clause": "ORDER BY",\
              "index_order_summary": {\
                "table": "`pre_forum_post`",\
                "index_provides_order": true,\
                "order_direction": "asc",\
                "index": "idx_dt",\
                "plan_changed": true,\
                "access_type": "index_scan"\
              } /* index_order_summary */\
            } /* reconsidering_access_paths_for_index_ordering */\

而在前面analyzing_range_alternativesconsidered_execution_plans阶段,都认为其他几个索引也是可选择的,直到这里才给强X了,你Y 看起来像是MySQL 5.6查询优化器的bug了,GOOGLE了一下,还真发有人已经反馈过类似的问题: MySQL bug 70245: incorrect costing for range scan causes optimizer to choose incorrect index

看完才发现,其实不是神马BUG,而是原来从5.6开始,增加了一个选项叫eq_range_index_dive_limit 的高级货,这货大概的用途是: 在较多等值查询(例如多值的IN查询)情景中,预估可能会扫描的记录数,从而选择相对更合适的索引,避免所谓的index dive问题。

当面临下面两种选择时:

1、索引代价较高,但结果较为精确;
2、索引代价较低,但结果可能不够精确;

简单说,选项 eq_range_index_dive_limit 的值设定了 IN列表中的条件个数上线,超过设定值时,会将执行计划分支从 1 变成 2

该值默认为10,但社区众多人反馈较低了,因此在5.7版本后,将默认值调整为200了。

不过,今天我们这里的案例却是想反的,因为优化器选择了看似代价低但精确的索引,实际却选择了更低效的索引。 因此,我们需要将其阈值调低,尝试设置 eq_range_index_dive_limit = 2 后(上面的例子中,IN条件里有2个值),再看下新的查询计划:

mysql> set eq_range_index_dive_limit = 2;

mysql> explain select * from pre_forum_post where tid=7932612 and `invisible` in('0','-2') order by dateline  limit 15\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: pre_forum_post
         type: range
possible_keys: PRIMARY,displayorder,first
          key: displayorder
      key_len: 4
          ref: NULL
         rows: 54
        Extra: Using index condition; Using filesort

卧槽,预估扫描记录数又降了557倍,相比最开始降了接近32万倍! 在这个案例中,虽然通过修改选项 eq_range_index_dive_limit 的阈值可以达到优化效果,但事实上更靠谱的做法是:直接删除 idx_dt 索引。 是的,没错,删除这个垃圾重复索引,因为实际上这个索引的用处不大,够坑爹吧~~

参考资料:
http://blog.163.com/li_hx/blog/static/18399141320147521735442/
http://mysqlserverteam.com/you-asked-for-it-new-default-for-eq_range_index_dive_limit/ https://www.facebook.com/note.php?note_id=10151533648715933
http://bugs.mysql.com/bug.php?id=70586
http://bugs.mysql.com/bug.php?id=67980
http://bugs.mysql.com/bug.php?id=70331

发表评论

电子邮件地址不会被公开。

To create code blocks or other preformatted text, indent by four spaces:

    This will be displayed in a monospaced font. The first four 
    spaces will be stripped off, but all other whitespace
    will be preserved.
    
    Markdown is turned off in code blocks:
     [This is not a link](http://example.com)

To create not a block, but an inline code span, use backticks:

Here is some inline `code`.

For more help see http://daringfireball.net/projects/markdown/syntax

Time limit is exhausted. Please reload CAPTCHA.