摘要:GaussDB(DWS) ROLLUP,CUBE,GROUPING SETS等OLAP函数的原理解析。
本文分享自华为云社区《GaussDB(DWS) OLAP函数浅析》,作者: DWS_Jack_2。
在一些报表场景中,经常会对数据做分组统计(group by),例如对一级部门下辖的二级部门员工数进行统计:
create table emp( id int, --工号 name text, --员工名 dep_1 text, --一级部门 dep_2 text --二级部门 ); gaussdb=# select count(*), dep_2 from emp group by dep_2; count | dep_2 -------+------- 200 | SRE 100 | EI (2 rows)
常见的统计报表业务中,通常需要进一步计算一级部门的“合计”人数,也就是二级部门各分组的累加,就可以借助于rollup,如下所示,比前面的分组计算结果多了一行合计的数据:
gaussdb=# select count(*), dep_2 from emp group by rollup(dep_2); count | dep_2 -------+------- 200 | SRE 100 | EI 300 | (3 rows)
如上是一种group by扩展的高级分组函数使用场景,这一类分组函数统称为OLAP函数,在GaussDB(DWS)中支持 ROLLUP,CUBE,GROUPING SETS,下面对这几种OLAP函数的原理和应用场景做一下分析。
首先我们来创建一张表,customer,用户信息表,其中包含了用户id,用户名,年龄,国家,用户级别,性别,余额等信息:
create table customer ( c_id char(16) not null, c_name char(20) , c_age integer , c_country varchar(20) , c_class char(10), c_sex text, c_balance numeric ); insert into customer values(1, 'tom', '20', 'China', '1', 'male', 300); insert into customer values(2, 'jack', '30', 'USA', '1', 'male', 100); insert into customer values(3, 'rose', '40', 'UK', '1', 'female', 200); insert into customer values(4, 'Frank', '60', 'GER', '1', 'male', 100); insert into customer values(5, 'Leon', '20', 'China', '2', 'male', 200); insert into customer values(6, 'Lucy', '20', 'China', '1', 'female', 500);
ROLLUP
本文开头的示例已经解释了,ROLLUP是在分组计算基础上增加了合计,从字面意思理解,就是从最小聚合级开始,聚合单位逐渐扩大,例如如下语句:
select c_country, c_class, sum(c_balance) from customer group by rollup(c_country, c_class) order by 1,2,3; c_country | c_class | sum -----------+------------+------ China | 1 | 800 China | 2 | 200 China | | 1000 GER | 1 | 100 GER | | 100 UK | 1 | 200 UK | | 200 USA | 1 | 100 USA | | 100 | | 1400 (10 rows)
该语句功能等价于如下:
select c_country, c_class, sum(c_balance) from customer group by c_country, c_class union all select c_country, null, sum(c_balance) from customer group by c_country union all select null, null, sum(c_balance) from customer order by 1,2,3; c_country | c_class | sum -----------+------------+------ China | 1 | 800 China | 2 | 200 China | | 1000 GER | 1 | 100 GER | | 100 UK | 1 | 200 UK | | 200 USA | 1 | 100 USA | | 100 | | 1400 (10 rows)
尝试理解一下
GROUP BY ROLLUP(A,B):
首先对(A,B)进行GROUP BY,然后对(A)进行GROUP BY,最后对全表进行GROUP BY操作
CUBE
CUBE从字面意思理解,就是各个维度的意思,也就是说全部组合,即聚合键中所有字段的组合的分组统计结果,例如如下语句:
select c_country, c_class, sum(c_balance) from customer group by cube(c_country, c_class) order by 1,2,3; c_country | c_class | sum -----------+------------+------ China | 1 | 800 China | 2 | 200 China | | 1000 GER | 1 | 100 GER | | 100 UK | 1 | 200 UK | | 200 USA | 1 | 100 USA | | 100 | 1 | 1200 | 2 | 200 | | 1400 (12 rows)
该语句功能等价于如下:
select c_country, c_class, sum(c_balance) from customer group by c_country, c_class union all select c_country, null, sum(c_balance) from customer group by c_country union all select null, null, sum(c_balance) from customer union all select NULL, c_class, sum(c_balance) from customer group by c_class order by 1,2,3; c_country | c_class | sum -----------+------------+------ China | 1 | 800 China | 2 | 200 China | | 1000 GER | 1 | 100 GER | | 100 UK | 1 | 200 UK | | 200 USA | 1 | 100 USA | | 100 | 1 | 1200 | 2 | 200 | | 1400 (12 rows)
理解一下
GROUP BY CUBE(A,B):
首先对(A,B)进行GROUP BY,然后依次对(A)、(B)进行GROUP BY,最后对全表进行GROUP BY操作。
GROUPING SETS
GROUPING SETS区别于ROLLUP和CUBE,并没有总体的合计功能,相当于从ROLLUP和CUBE的结果中提取出部分记录,例如如下语句:
select c_country, c_class, sum(c_balance) from customer group by grouping sets(c_country, c_class) order by 1,2,3; c_country | c_class | sum -----------+------------+------ China | | 1000 GER | | 100 UK | | 200 USA | | 100 | 1 | 1200 | 2 | 200 (6 rows)
该语句功能等价于如下:
select c_country, null, sum(c_balance) from customer group by c_country union all select null, c_class, sum(c_balance) from customer group by c_class order by 1,2,3; c_country | ?column? | sum -----------+------------+------ China | | 1000 GER | | 100 UK | | 200 USA | | 100 | 1 | 1200 | 2 | 200 (6 rows)
理解一下
GROUP BY GROUPING SETS(A,B):
分别对(B)、(A)进行GROUP BY计算
目前在GaussDB(DWS)中,OLAP函数的实现,会有排序(sort)操作,相比等价的union all操作,效率并不会有提升,后续会通过mixagg的支持来提升OLAP函数的执行效率,有兴趣的同学,可以explain打印一下计划,来看一下OLAP函数的执行流程。