优化案例 | CASE WHEN进行SQL改写优化

待优化场景
发现SLOW QUERY LOG中有下面这样一条记录:mysql

...
# Query_time: 59.503827  Lock_time: 0.000198  Rows_sent: 641227  Rows_examined: 13442472  Rows_affected: 0
...
select uid,sum(power) powerup from t1 where 
date>='2017-03-31' and 
UNIX_TIMESTAMP(STR_TO_DATE(concat(date,' ',hour),'%Y-%m-%d %H'))>=1490965200 and 
UNIX_TIMESTAMP(STR_TO_DATE(concat(date,' ',hour),'%Y-%m-%d %H'))<1492174801  and 
aType in (1,6,9) group by uid;

实话说,看到这个SQL我也忍不住想骂人啊,到底是哪一个脑残的XX狗设计的?sql

居然把日期时间中的 date 和 hour 给独立出来成两列,查询时再合并成一个新的条件,简直无力吐槽。函数

吐槽归吐槽,该干活还得干活,谁让咱是DBA呢,SQL优化是咱的拿手好戏不是嘛~性能

SQL优化之路
一、SQL优化思路
不厌其烦地再说一遍SQL优化思路。优化

想要优化一个SQL,通常来讲就是先看执行计划,观察是否尽量用到索引,ui

同时要关注预计扫描的行数,插件

以及是否产生了临时表(Using temporary) 或者 设计

是否须要进行排序(Using filesort),code

想办法消除这些状况。排序

二、SQL性能瓶颈定位
毫无疑问,想要优化,先看表DDL以及执行计划:

CREATE TABLE `t1` (
  `id` bigint(20) unsigned NOT NULL AUTO_INCREMENT,
  `date` date NOT NULL DEFAULT '0000-00-00',
  `hour` char(2) NOT NULL DEFAULT '00',
  `kid` int(4) NOT NULL DEFAULT '0',
  `uid` int(11) NOT NULL DEFAULT '0',
  `aType` tinyint(2) NOT NULL DEFAULT '0',
  `src` tinyint(2) NOT NULL DEFAULT '1',
  `aid` int(11) NOT NULL DEFAULT '1',
  `acount` int(11) NOT NULL DEFAULT '1',
  `power` decimal(20,2) DEFAULT '0.00',
  PRIMARY KEY (`id`,`date`),
  UNIQUE KEY `did` (`date`,`hour`,`kid`,`uid`,`aType`,`src`,`aid`)
) ENGINE=InnoDB AUTO_INCREMENT=50486620 DEFAULT CHARSET=utf8mb4
/*!50500 PARTITION BY RANGE  COLUMNS(`date`)
(PARTITION p20170316 VALUES LESS THAN ('2017-03-17') ENGINE = InnoDB,
 PARTITION p20170317 VALUES LESS THAN ('2017-03-18') ENGINE = InnoDB
...

yejr@imysql.com[myDB]> EXPLAIN select uid,sum(power) powerup from t1 where 
date>='2017-03-31' and 
UNIX_TIMESTAMP(STR_TO_DATE(concat(date,' ',hour),'%Y-%m-%d %H'))>=1490965200 and 
UNIX_TIMESTAMP(STR_TO_DATE(concat(date,' ',hour),'%Y-%m-%d %H'))<1492174801  and 
aType in (1,6,9) group by uid\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: t1
   partitions: p20170324,p20170325,....all partition
         type: ALL
possible_keys: did
          key: NULL
      key_len: NULL
          ref: NULL
         rows: 25005577
     filtered: 15.00
        Extra: Using where; Using temporary; Using filesort

明显的,这个SQL效率很是低,全表扫描、没有索引、有临时表、须要额外排序,什么倒霉催的全遇上了。

三、优化思考
这个SQL是想统计符合条件的power列总和,虽然 date 列已有索引,但WHERE子句中却对 date 列加了函数,并且仍是 date 和 hour 两列的组合条件,那就没法用到这个索引了。

还好,有个聪明伶俐的妹子,突发起想(事实上这位妹子原本就擅长作SQL优化的~),能够用 CASE WHEN 方法来改造下SQL,改为像下面这样的:

select uid,sum(powerup+powerup1) from
(
   select uid,
          case when concat(date,' ',hour) >='2017-03-24 13:00' then power else '0' end as powerup,
          case when concat(date,' ',hour) < '2017-03-25 13:00' then power else '0' end as powerup1
   from t1
   where date>='2017-03-24' 
   and   date <'2017-03-25'
   and  aType in (1,6,9)
) a  group by uid;

是否是颇有才,直接把这个没办法用到索引的条件给用CASE WHEN来改造了。看看新的SQL执行计划:

*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: t1
   partitions: p20170324
         type: range
possible_keys: did
          key: idx2_date_addRedType
      key_len: 4
          ref: NULL
         rows: 876375
     filtered: 30.00
        Extra: Using index condition; Using temporary; Using filesort

看看这个SQL的执行代价:

+----------------------------+---------+
| Variable_name              | Value   |
+----------------------------+---------+
| Handler_read_first         | 1       |
| Handler_read_key           | 1834590 |
| Handler_read_last          | 0       |
| Handler_read_next          | 1834589 |
| Handler_read_prev          | 0       |
| Handler_read_rnd           | 232276  |
| Handler_read_rnd_next      | 232277  |
+----------------------------+---------+

及其SLOW QUERY LOG记录的信息:

# Query_time: 6.381254  Lock_time: 0.000166  Rows_sent: 232276  Rows_examined: 2299141  Rows_affected: 0
# Bytes_sent: 4237347  Tmp_tables: 1  Tmp_disk_tables: 0  Tmp_table_sizes: 4187168
# InnoDB_trx_id: 0
# QC_Hit: No  Full_scan: No  Full_join: No  Tmp_table: Yes  Tmp_table_on_disk: No
# Filesort: Yes  Filesort_on_disk: No  Merge_passes: 0
#   InnoDB_IO_r_ops: 0  InnoDB_IO_r_bytes: 0  InnoDB_IO_r_wait: 0.000000
#   InnoDB_rec_lock_wait: 0.000000  InnoDB_queue_wait: 0.000000
#   InnoDB_pages_distinct: 9311

看起来还不是太理想啊,虽然再也不扫描全表了,但毕竟仍是 有临时表 和 额外排序,想办法消除后再对比看下。

有个变化不知道你们注意到没,新的SLOW QUERY LOG记录多了很多信息,这是由于用了Percona分支版本的插件才支持,这个功能确实不错,甚至还能记录Profiling的详细信息,强烈推荐。

咱们新建个 uid 列上的索引,看看能除临时表及排序后的代价如何,看看这个的开销会不会更低。

yejr@imysql.com[myDB]> ALTER TABLE t1 ADD INDEX idx_uid(uid);
yejr@imysql.com[myDB]> EXPLAIN select uid,sum(powerup+powerup1) from
(
   select uid,
          case when concat(date,' ',hour) >='2017-03-24 13:00' then power else '0' end as powerup,
          case when concat(date,' ',hour) < '2017-03-25 13:00' then power else '0' end as powerup1
   from t1
   where date>='2017-03-24' 
   and   date <'2017-03-25'
   and  aType in (1,6,9)
) a  group by uid\G

*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: if_date_hour_army_count
   partitions: p20170331,p20170401...
         type: index
possible_keys: did,idx_uid
          key: idx_uid
      key_len: 4
          ref: NULL
         rows: 12701520
     filtered: 15.00
        Extra: Using where

看看添加索引后SQL的执行代价:

+----------------------------+---------+
| Variable_name              | Value   |
+----------------------------+---------+
| Handler_read_first         | 1       |
| Handler_read_key           | 1       |
| Handler_read_last          | 0       |
| Handler_read_next          | 1834589 |
| Handler_read_prev          | 0       |
| Handler_read_rnd           | 0       |
| Handler_read_rnd_next      | 0       |
+----------------------------+---------+

及其SLOW QUERY LOG记录的信息:

# Query_time: 5.772286  Lock_time: 0.000330  Rows_sent: 232276  Rows_examined: 1834589  Rows_affected: 0
# Bytes_sent: 4215071  Tmp_tables: 0  Tmp_disk_tables: 0  Tmp_table_sizes: 0
# InnoDB_trx_id: 0
# QC_Hit: No  Full_scan: Yes  Full_join: No  Tmp_table: No  Tmp_table_on_disk: No
# Filesort: No  Filesort_on_disk: No  Merge_passes: 0
#   InnoDB_IO_r_ops: 0  InnoDB_IO_r_bytes: 0  InnoDB_IO_r_wait: 0.000000
#   InnoDB_rec_lock_wait: 0.000000  InnoDB_queue_wait: 0.000000
#   InnoDB_pages_distinct: 11470

咱们注意到,虽然加了 uid 列索引后的SQL扫描的data page更多了,但执行效率实际上是更高的,由于消除了 临时表 和 额外排序,这从 Handlerread% 的结果中也能看出来,很显然它的顺序I/O更多,随机I/O更少,因此虽然须要扫描的 data page 更多,实际上效率倒是更快的。

后记再想一想这个SQL还有优化空间吗,显然是有的,那就是把数据表从新设计,将date和hour列整合到一块儿,这样就不用费劲的拼凑条件而且也能用到索引了。

相关文章
相关标签/搜索