----原语句（运行18min）
SELECT
bb.ip
FROM
(
SELECT
ip ,
sum(click) click_num,
round(sum(click) / sum(imp), 4) user_click_rate
FROM
schema.srctable1
WHERE
date = '20171020'
AND ip IS NOT NULL
AND imp > 0
GROUP BY
ip
)
bb
LEFT OUTER JOIN
(
SELECT
round(sum(click) / sum(imp), 4) avg_click_rate
FROM
schema.srctable1
WHERE
date = '20171020'
)
aa
LEFT OUTER JOIN schema.dstable cc
on
cc.ip = bb.ip
WHERE
cc.ip is null
AND
(
bb.user_click_rate > aa.avg_click_rate * 3
AND click_num > 500
)
OR
(
click_num > 1000
)
分析：
1、aa表存放的就是一个指标数据,1条记录,列为小表2、bb表存放的是按ip聚合的明细数据,记录很多,列为大表
3、cc表用来过滤ip,数量也很小,列为过滤表,作用很小。
查看执行计划,发现bb与aa进行left outer join时，引发了shuffle过程,造成大量的磁盘及网络IO,影响性能。
解决策略
优化方案1:调整大小表位置,将小表放在左边后，提升至29s (该方案一直不太明白为啥会提升，执行计划里显示的也就是大小表位置调换下而已，跟之前的没其他区别)优化方案2: 将 or 改成 union,提升至35s(各种调整,一直怀疑跟or有关系,后面调整成union其他不变,果真效率不一样;但方案1只是调整了下大小表顺序,并未调整其他,其效率同样提升很大;不太明白sparksql内部到底走了什么优化机制,后面继续研究);
优化方案3： 采用cache+broadcast方式,提升至20s（该方案将小表缓存至内存，进行map侧关联）
方案具体实施
----方案2:or 改成 union（运行35s）
I select
aa.ip
from
(
SELECT
bb.ip ip
FROM
(
SELECT
ip ,
sum(click) click_num,
round(sum(click) / sum(imp), 4)
user_click_rate
FROM
schema.srctable1
WHERE
date = '20171020'
AND ip IS NOT NULL
AND imp > 0
GROUP BY
ip
)
bb
LEFT OUTER JOIN
(
SELECT
round(sum(click) / sum(imp), 4)
avg_click_rate
FROM
schema.srctable1
WHERE
date = '20171020'
)
aa
WHERE
(
bb.user_click_rate > aa.avg_click_rate * 3
AND click_num > 20
)
union
SELECT
bb.ip ip
FROM
(
SELECT
ip ,
sum(click) click_num,
round(sum(click) / sum(imp), 4)
user_click_rate
FROM
schema.srctable1
WHERE
date = '20171020'
AND ip IS NOT NULL
AND imp > 0
GROUP BY
ip
)
bb
LEFT OUTER JOIN
(
SELECT
round(sum(click) / sum(imp), 4)
avg_click_rate
FROM
schema.srctable1
WHERE
date = '20171020'
)
aa
WHERE
click_num > 40
)
aa
LEFT OUTER JOIN schema.dstable cc
on
aa.ip = cc.ip
where
cc.ip is null
-----cache+broadcast方式（20s）
原理：使用broadcast将会把小表分发到每台执行节点上，因此，关联操作都在本地完成，基本就取消了shuffle的过程，运行效率大幅度提高。
cache table cta
as
SELECT
round(sum(click) / sum(imp), 4) avg_click_rate
FROM
schema.srctable1
WHERE
date = '20171020';
INSERT into TABLE schema.dstable
SELECT
bb.ip
FROM
(
SELECT
ip ,
sum(click) click_num,
round(sum(click) / sum(imp), 4) user_click_rate
FROM
schema.srctable1
WHERE
date = '20171020'
AND ip IS NOT NULL
AND imp > 0
GROUP BY
ip
)
bb
LEFT OUTER JOIN cta aa
LEFT OUTER JOIN schema.dstable cc
on
cc.ip = bb.ip
WHERE
cc.ip is null
AND
(
bb.user_click_rate > aa.avg_click_rate * 3
AND click_num > 500
)
OR
(
click_num > 1000
)
注意：
cache 表不一定会被广播到Executor，执行map side join，还受另外一个参数：spark.sql.autoBroadcastJoinThreshold影响，该参数判断是否将该表广播；
spark.sql.autoBroadcastJoinThreshold参数默认值是10M，所以只有cache的表小于10M的才被广播到Executor上去执行map side join。