还在用变量去实现多维度分组排序吗?你 out 了!
- 2022 年 7 月 11 日
本文字数:8149 字
阅读完需:约 27 分钟
作者: Gin 原文来源:https://tidb.net/blog/07531acc
目录 一、什么是窗口函数 二、用窗口函数实现分组内排序 三、基于窗口函数的高效分页批处理方案
一、什么是窗口函数
窗口函数(Window Function)又叫开窗函数,是一种常见的 OLAP 函数,与聚合函数不同,窗口函数可以按多个维度分别做排序,简化了复杂分析场景的 SQL 逻辑。常见的单机数据库一般都支持窗口函数,TiDB v3.0,MySQL 8.0 版本也开始支持窗口函数功能。
二、用窗口函数实现分组内排序
分组并对组内排序是使用窗口函数的常见场景。
首先我们制作一张学生成绩表,包含学生姓名,学号,科目,以及科目成绩字段,并写入一些数据:
mysql> select * from class_score;
+--------------+-----------+-------------------------+-----------+
| stuname | stuno | course | courscore |
+--------------+-----------+-------------------------+-----------+
| SpongeBob | 201903001 | LinearAlgebra | 60.5 |
| SpongeBob | 201903001 | AdvancedMathematics | 55.0 |
| SpongeBob | 201903001 | Physics | 65.0 |
| SpongeBob | 201903001 | ProbabilityTheory | 87.0 |
| SpongeBob | 201903001 | PrincipleofStatistics | 90.0 |
| SpongeBob | 201903001 | OperatingSystem | 95.0 |
| SpongeBob | 201903001 | FundamentalsofCompiling | 43.0 |
| SpongeBob | 201903001 | DiscreteMathematics | 72.0 |
| SpongeBob | 201903001 | PrinciplesofDatabase | 88.0 |
| SpongeBob | 201903001 | English | 79.0 |
| SpongeBob | 201903001 | OpBasketball | 92.0 |
| SpongeBob | 201903001 | OpTennis | 94.0 |
| PatrickStar | 201903011 | LinearAlgebra | 6.5 |
| PatrickStar | 201903011 | AdvancedMathematics | 5.0 |
| PatrickStar | 201903011 | Physics | 6.0 |
| PatrickStar | 201903011 | ProbabilityTheory | 12.0 |
| PatrickStar | 201903011 | PrincipleofStatistics | 20.0 |
| PatrickStar | 201903011 | OperatingSystem | 36.0 |
| PatrickStar | 201903011 | FundamentalsofCompiling | 2.0 |
| PatrickStar | 201903011 | DiscreteMathematics | 14.0 |
| PatrickStar | 201903011 | PrinciplesofDatabase | 9.0 |
| PatrickStar | 201903011 | English | 60.0 |
| PatrickStar | 201903011 | OpTableTennis | 12.0 |
| PatrickStar | 201903011 | OpPiano | 99.0 |
| MonkeyDLuffy | 201803015 | LinearAlgebra | 92.5 |
| MonkeyDLuffy | 201803015 | AdvancedMathematics | 95.5 |
| MonkeyDLuffy | 201803015 | Physics | 63.5 |
| MonkeyDLuffy | 201803015 | ProbabilityTheory | 76.0 |
| MonkeyDLuffy | 201803015 | PrincipleofStatistics | 69.0 |
| MonkeyDLuffy | 201803015 | OperatingSystem | 90.5 |
| MonkeyDLuffy | 201803015 | FundamentalsofCompiling | 88.0 |
| MonkeyDLuffy | 201803015 | DiscreteMathematics | 89.0 |
| MonkeyDLuffy | 201803015 | PrinciplesofDatabase | 60.5 |
| MonkeyDLuffy | 201803015 | English | 43.0 |
| MonkeyDLuffy | 201803015 | OpSwimming | 67.0 |
| MonkeyDLuffy | 201803015 | OpFencing | 76.0 |
+--------------+-----------+-------------------------+-----------+
36 rows in set (0.01 sec)
业务需求 1:计算出每科成绩的前两名的姓名、学号和成绩
这是一个难以用聚合函数实现的需求,由于长期不支持窗口函数,MySQL 社区普遍推荐使用用户变量的方式来实现,具体实现方式如下:
mysql> SET @z := NULL;
Query OK, 0 rows affected (0.00 sec)
mysql> SET @ROW_NUM := 0;
Query OK, 0 rows affected (0.00 sec)
mysql> select course, stuname, stuno, courscore from (select course, @ROW_NUM := IF(course = @z, @ROW_NUM + 1, 1) as ROW_NUM, @z := course AS z, stuname, courscore, stuno FROM (select * from class_score order by course, courscore desc) t1) t2 where t2.ROW_NUM<=2;
+-------------------------+--------------+-----------+-----------+
| course | stuname | stuno | courscore |
+-------------------------+--------------+-----------+-----------+
| AdvancedMathematics | MonkeyDLuffy | 201803015 | 95.5 |
| AdvancedMathematics | SpongeBob | 201903001 | 55.0 |
| DiscreteMathematics | MonkeyDLuffy | 201803015 | 89.0 |
| DiscreteMathematics | SpongeBob | 201903001 | 72.0 |
| English | SpongeBob | 201903001 | 79.0 |
| English | PatrickStar | 201903011 | 60.0 |
| FundamentalsofCompiling | MonkeyDLuffy | 201803015 | 88.0 |
| FundamentalsofCompiling | SpongeBob | 201903001 | 43.0 |
| LinearAlgebra | MonkeyDLuffy | 201803015 | 92.5 |
| LinearAlgebra | SpongeBob | 201903001 | 60.5 |
| OpBasketball | SpongeBob | 201903001 | 92.0 |
| OpFencing | MonkeyDLuffy | 201803015 | 76.0 |
| OpPiano | PatrickStar | 201903011 | 99.0 |
| OpSwimming | MonkeyDLuffy | 201803015 | 67.0 |
| OpTableTennis | PatrickStar | 201903011 | 12.0 |
| OpTennis | SpongeBob | 201903001 | 94.0 |
| OperatingSystem | SpongeBob | 201903001 | 95.0 |
| OperatingSystem | MonkeyDLuffy | 201803015 | 90.5 |
| Physics | SpongeBob | 201903001 | 65.0 |
| Physics | MonkeyDLuffy | 201803015 | 63.5 |
| PrincipleofStatistics | SpongeBob | 201903001 | 90.0 |
| PrincipleofStatistics | MonkeyDLuffy | 201803015 | 69.0 |
| PrinciplesofDatabase | SpongeBob | 201903001 | 88.0 |
| PrinciplesofDatabase | MonkeyDLuffy | 201803015 | 60.5 |
| ProbabilityTheory | SpongeBob | 201903001 | 87.0 |
| ProbabilityTheory | MonkeyDLuffy | 201803015 | 76.0 |
+-------------------------+--------------+-----------+-----------+
26 rows in set (0.01 sec)
通过定义两个用户变量,一个用于切换到下一组,另一个用来发放行号,以此来通过嵌套循环的方式来实现为每组单独发放行号。缺点是不能处理相同分数名次并列的情况,并且嵌套太多,逻辑比较复杂,每次计算都要为变量重新赋值。
来看一下窗口函数的实现方式,仅需要一条 SQL,一个子查询就可以得出各科成绩的前两名,注意这里使用的 rank() 函数可以识别相同分数名次并列的情况,也就是说假如一科出现了两人并列第一,使用下面的 SQL 可以公平的把并列第一的情况展现出来,这是用户变量难以实现的。
mysql> select course, stuname, stuno, courscore from (select *, rank() over(partition by course order by course, courscore desc) as RANK_ from class_score) t where t.RANK_<=2;
+-------------------------+--------------+-----------+-----------+
| course | stuname | stuno | courscore |
+-------------------------+--------------+-----------+-----------+
| AdvancedMathematics | MonkeyDLuffy | 201803015 | 95.5 |
| AdvancedMathematics | SpongeBob | 201903001 | 55.0 |
| DiscreteMathematics | MonkeyDLuffy | 201803015 | 89.0 |
| DiscreteMathematics | SpongeBob | 201903001 | 72.0 |
| English | SpongeBob | 201903001 | 79.0 |
| English | PatrickStar | 201903011 | 60.0 |
| FundamentalsofCompiling | MonkeyDLuffy | 201803015 | 88.0 |
| FundamentalsofCompiling | SpongeBob | 201903001 | 43.0 |
| LinearAlgebra | MonkeyDLuffy | 201803015 | 92.5 |
| LinearAlgebra | SpongeBob | 201903001 | 60.5 |
| OpBasketball | SpongeBob | 201903001 | 92.0 |
| OpFencing | MonkeyDLuffy | 201803015 | 76.0 |
| OpPiano | PatrickStar | 201903011 | 99.0 |
| OpSwimming | MonkeyDLuffy | 201803015 | 67.0 |
| OpTableTennis | PatrickStar | 201903011 | 12.0 |
| OpTennis | SpongeBob | 201903001 | 94.0 |
| OperatingSystem | SpongeBob | 201903001 | 95.0 |
| OperatingSystem | MonkeyDLuffy | 201803015 | 90.5 |
| Physics | SpongeBob | 201903001 | 65.0 |
| Physics | MonkeyDLuffy | 201803015 | 63.5 |
| PrincipleofStatistics | SpongeBob | 201903001 | 90.0 |
| PrincipleofStatistics | MonkeyDLuffy | 201803015 | 69.0 |
| PrinciplesofDatabase | SpongeBob | 201903001 | 88.0 |
| PrinciplesofDatabase | MonkeyDLuffy | 201803015 | 60.5 |
| ProbabilityTheory | SpongeBob | 201903001 | 87.0 |
| ProbabilityTheory | MonkeyDLuffy | 201803015 | 76.0 |
+-------------------------+--------------+-----------+-----------+
26 rows in set (0.01 sec)
业务需求 2:计算出每科成绩第一名与第二名之间的分差
TiDB 提供 lead() 与 lag() 函数来获取组内数据排序后的下一行或上一行的列值,此处正是使用了 lead() 函数来获取下一行的列值,通过子查询的方式即可计算出第一名与第二名之间的分差:
mysql> select course, courscore, courscore - lead_ as delta from (select *, lead(courscore,1) over(partition by course order by course, courscore desc) as lead_, rank() over(partition by course order by course, courscore desc) as RANK_ from class_score) t where t.RANK_=1;
+-------------------------+-----------+-------+
| course | courscore | delta |
+-------------------------+-----------+-------+
| AdvancedMathematics | 95.5 | 40.5 |
| DiscreteMathematics | 89.0 | 17.0 |
| English | 79.0 | 19.0 |
| FundamentalsofCompiling | 88.0 | 45.0 |
| LinearAlgebra | 92.5 | 32.0 |
| OpBasketball | 92.0 | NULL |
| OpFencing | 76.0 | NULL |
| OpPiano | 99.0 | NULL |
| OpSwimming | 67.0 | NULL |
| OpTableTennis | 12.0 | NULL |
| OpTennis | 94.0 | NULL |
| OperatingSystem | 95.0 | 4.5 |
| Physics | 65.0 | 1.5 |
| PrincipleofStatistics | 90.0 | 21.0 |
| PrinciplesofDatabase | 88.0 | 27.5 |
| ProbabilityTheory | 87.0 | 11.0 |
+-------------------------+-----------+-------+
16 rows in set (0.00 sec)
三、基于窗口函数的高效分页批处理方案
窗口函数作为数据库的高级分析功能,它的应用场景不仅限于分组内排序,我们还可以利用窗口函数做很多有意思的事情,比如本案例用窗口函数来大幅优化跑批中的分页处理效率。
我们用 sysbench 创建一张表并加载一些数据,用这张表来模拟批量处理逻辑。
首先初始化一张表 sbtest1,其表结构如下,其中 id 字段为整型主键:
mysql> desc sbtest1;
+-------+-----------+------+------+---------+----------------+
| Field | Type | Null | Key | Default | Extra |
+-------+-----------+------+------+---------+----------------+
| id | int(11) | NO | PRI | NULL | auto_increment |
| k | int(11) | NO | MUL | 0 | |
| c | char(120) | NO | | | |
| pad | char(60) | NO | | | |
+-------+-----------+------+------+---------+----------------+
4 rows in set (0.00 sec)
初始化时加载了 100 万行数据,之后我们删除掉其中一部分,通过这样的方式使 id 值不再连续,弱化分页时对于 id 值的依赖。当前表中剩余数据有 90 万行左右:
mysql> select count(*) from sbtest1;
+----------+
| count(*) |
+----------+
| 899997 |
+----------+
1 row in set (0.65 sec)
表内数据预览:
mysql> select * from sbtest1 limit 6;
+--------+--------+-------------------------------------------------------------------------------------------------------------------------+-------------------------------------------------------------+
| id | k | c | pad |
+--------+--------+-------------------------------------------------------------------------------------------------------------------------+-------------------------------------------------------------+
| 170713 | 502585 | 68207710198-92682096687-30191949979-36606876762-68131108662-05227395575-42775011851-25226186240-86628605904-92905646658 | 92965159868-07234410731-39167064470-14286085716-15715308680 |
| 170715 | 594870 | 03482720054-50379763215-87903836122-97559417898-49419423256-08561919665-14395666373-04552411341-51225532045-80056729812 | 14534783486-12748024297-66217900494-07062661389-59419864770 |
| 170716 | 618106 | 17284178744-35252021030-57793972189-12648949390-90678614158-50453793363-79361198568-92739087625-90147799094-56275382145 | 96022213702-57054390589-17717245768-83668730988-26655128451 |
| 170717 | 498071 | 55266913813-66118089063-10841700714-78346894223-87037025257-46356741961-50684103191-23859048041-87607902200-58092836685 | 85952977843-18323978167-65380568194-90178704467-17391816925 |
| 170718 | 500843 | 81176419361-91278769025-45575469479-70005546210-57581523030-24528178176-84655463505-48851510236-43885747093-01732211221 | 56651630364-99235825673-25852643818-33561663285-01699695675 |
| 170719 | 499063 | 87982690236-17188898588-98406118277-04805507744-90184035670-09591916010-78045349706-89374841792-79952082330-08177876709 | 16885918921-25441055158-88415348869-22003000705-82198521530 |
+--------+--------+-------------------------------------------------------------------------------------------------------------------------+-------------------------------------------------------------+
6 rows in set (0.01 sec)
常见的分页更新 SQL 一般使用主键 / 唯一索引进行排序,以确保相邻的两页之间没有空隙或重叠,配合 MySQL limit 语法中非常好用的 offset 功能来按固定行数拆分页面,拆分后的页面被包装在独立的事务中,可以灵活的进行逐页或批量对数据进行更新。
begin;
update sbtest1 set pad='new_value' where id in (select id from sbtest1 order by id limit 0,10000);
commit;
begin;
update sbtest1 set pad='new_value' where id in (select id from sbtest1 order by id limit 10000,10000);
commit;
begin;
update sbtest1 set pad='new_value' where id in (select id from sbtest1 order by id limit 20000,10000);
commit;
这种方案逻辑清晰,SQL 易于编写,但它有着明显的劣势,由于需要对主键 / 唯一索引进行排序,越靠后的页面需要参与排序的行数越多,TiKV 中扫描数据的压力也越大,批量整体处理效率就越低,当批量的整体数据量比较大时,很可能会占用过多计算资源,甚至触发性能瓶颈,影响联机业务。
下面案例是一种改进方案,通过灵活运用窗口函数 row_number() 将数据按照主键排序后赋予行号,再通过聚合函数按照设置好的页面大小对行号进行分组,以计算出每页的最大值和最小值。
mysql> select min(t.id) as start_key, max(t.id) as end_key, count(*) as page_size from (select *, row_number() over(order by id) as row_num from sbtest1) t group by floor((t.row_num-1)/50000) order by start_key;
+-----------+---------+-----------+
| start_key | end_key | page_size |
+-----------+---------+-----------+
| 1 | 55556 | 50000 |
| 55557 | 111111 | 50000 |
| 111112 | 166667 | 50000 |
| 166668 | 222222 | 50000 |
| 222223 | 277778 | 50000 |
| 277779 | 333333 | 50000 |
| 333335 | 388889 | 50000 |
| 388890 | 444445 | 50000 |
| 444446 | 500000 | 50000 |
| 500001 | 555556 | 50000 |
| 555557 | 611111 | 50000 |
| 611112 | 666667 | 50000 |
| 666668 | 722223 | 50000 |
| 722225 | 777779 | 50000 |
| 777780 | 833335 | 50000 |
| 833336 | 888891 | 50000 |
| 888892 | 944447 | 50000 |
| 944448 | 1000000 | 49997 |
+-----------+---------+-----------+
18 rows in set (1.87 sec)
将这个结果集作为批量处理的元信息,这样在批量处理阶段只需要使用 between...and...
来圈定好每个页面的数据,多个页面并发的进行批量更新即可,由于元信息的计算阶段使用主键 / 唯一索引进行排序,并用 row_number() 函数赋予了唯一序号,因此也可以避免在两个相邻的页面中出现空隙或重叠。
使用这种方案可以显著避免由于频繁,大量的排序造成的性能损耗,进而大幅提升批量处理的整体效率。
mysql> update sbtest1 set pad='new_value' where id between 1 and 55556;
Query OK, 50000 rows affected (3.51 sec)
Rows matched: 50000 Changed: 50000 Warnings: 0
mysql> update sbtest1 set pad='new_value' where id between 55557 and 111111;
Query OK, 50000 rows affected (2.14 sec)
Rows matched: 50000 Changed: 50000 Warnings: 0
mysql> update sbtest1 set pad='new_value' where id between 111112 and 166667;
Query OK, 50000 rows affected (2.21 sec)
Rows matched: 50000 Changed: 50000 Warnings: 0
四、复合主键分页案例
制作元信息表
操作分页的案例
另外可以使用隐藏字段 _tidb_rowid 做分页使用。
版权声明: 本文为 InfoQ 作者【TiDB 社区干货传送门】的原创文章。
原文链接:【http://xie.infoq.cn/article/ce0f3715e6ef6b901885a5740】。文章转载请联系作者。
TiDB 社区干货传送门
TiDB 社区官网:https://tidb.net/ 2021.12.15 加入
TiDB 社区干货传送门是由 TiDB 社区中布道师组委会自发组织的 TiDB 社区优质内容对外宣布的栏目,旨在加深 TiDBer 之间的交流和学习。一起构建有爱、互助、共创共建的 TiDB 社区 https://tidb.net/
评论