【大数据技巧】MaxCompute优化去重计算的性能

免费开通大数据服务：https://www.aliyun.com/product/odps

转载自dachuan

随着双十一数据量的暴增，之前用distinct去重可以简单处理的场景，现在消耗的时间成倍增长。如果用了multiple distinct，那就更要警惕，因为多重去重本身会带来数据量的成倍增长，很可能10分钟的任务，在双十一期间会跑上几个小时都没有结果。

这里介绍一个小技巧，其实在稳定性手册里面已经有过介绍，不过总感觉没有看懂。最近正好做了一次优化，于是在这里小结一下：

例如原来的代码是这样：

select D1,D2,
count(distinct case when A is not null then B end) as B_distinct_cnt 
from xxx group by D1,D2

那么优化方案可以是这样：

create table tmp1
as
select D1,D2,B,
count( case when A is not null then B end ) as B_cnt
from xxx
group by D1, D1, B


select D1,D2,
sum(case when B_cnt > 0 then 1 else 0 end) as B_distinct_cnt
from tmp1
group by D1,D2

多重去重的优化也可以采用上面的方案，只是要注意Group By的Key是以源表聚合维度为基础，根据distinct计算的值进行组合。

例如下面的这个例子：

select D1,D2,
count(distinct case when A is not null then B end) as B_distinct_cnt ,
count(distinct case when E is not null then C end) as C_distinct_cnt 
from xxx group by D1,D2

那么优化方案可以是：

create table tmp1
as
select D1,D2,B,
count( case when A is not null then B end ) as B_cnt
from xxx
group by D1, D1, B

create table tmp1_1
as
select D1,D2,
sum(case when B_cnt > 0 then 1 else 0 end) as B_distinct_cnt
from tmp1
group by D1,D2

create table tmp2
as
select D1,D2,C,
count( case when E is not null then C end ) as C_cnt
from xxx
group by D1, D1, C

create table tmp2_1
as
select D1,D2,
sum(case when C_cnt > 0 then 1 else 0 end) as C_distinct_cnt
from tmp1
group by D1,D2

select 
t1.D1,t1.D2,
t1.B_distinct_cnt,
t2.C_distinct_cnt
from tmp1_1 t1
left outer join tmp2_1 t2
on t1.D1=t2.D1 and t1.D2=t2.D2

欢迎加入 MaxCompute 钉钉群讨论

原文链接：

http://click.aliyun.com/m/13956/