sql优化实战：把full join改为left join +union all（从5分钟降为10秒）

今天收到一个需求，要改写一个报表的逻辑，当改完之后，再次运行，发现运行超时。

因为特殊原因，无法访问客户的服务器，没办法查看sql的执行计划、没办法知道表中的索引情况，所以，尝试从语句的改写上来优化。

一、原始语句如下：

select  isnull(vv.customer_id,v.customer_id) as customer_id,
isnull(vv.business_date,replace(v.business_date,'-','')) as business_date,
v.prod_id,
v.sales,
vv.visit_count,
v.all_sales
from
(
SELECT  a.customer_id ,
max(month)+'-01' as business_date,
a.PROD_ID ,
SUM(CAST(VALUE AS NUMERIC(38, 3))) sales,
sum(SUM(CAST(VALUE AS NUMERIC(38, 3)))) over(partition by a.customer_id) as all_sales

FROM    TB_IMPORT_SALES a
WHERE   a.customer_id IS NOT NULL
AND a.PROD_ID IS NOT NULL
and a.month='2016-11'
GROUP BY a.customer_id ,
a.PROD_ID
)v
full join
(
SELECT customer_id,
max(a.business_date) as business_date,
COUNT(*) AS VISIT_COUNT
FROM TB_CALL_STORE a WITH(NOLOCK)
inner join TB_TIME d
on a.business_date = d.t_date
where d.section ='2016-11'
GROUP BY customer_id
)vv
on v.customer_id = vv.customer_id

原来是left join，虽然查询比较慢，但是2分钟能查出来，现在按照业务要求，需要看到所有数据，所以改成了full join，改了之后5分钟都查不出结果。

二、改写后的代码

select  v.customer_id,
replace(max(v.business_date),'-','') as business_date,
v.prod_id,
max(v.sales_volume) sales_volume ,
max(v.visit_count) visit_count,
max(v.all_sales_volume) all_sales_volume
from
(
SELECT  a.customer_id ,
max(biz_month)+'-01' as business_date,
a.PROD_ID ,
SUM(CAST(VALUE1 AS NUMERIC(38, 8))) sales_volume,
sum(SUM(CAST(VALUE1 AS NUMERIC(38, 8)))) over(partition by a.customer_id) as all_sales_volume,
null as visit_count

FROM    TB_IMPORT_SALES a
WHERE   a.customer_id IS NOT NULL
AND a.PROD_ID IS NOT NULL
and a.month='2016-11'
GROUP BY a.customer_id ,
a.PROD_ID
union all

SELECT customer_id,
max(a.business_date) as business_date,
p.prod_id,
null,
null,
COUNT(*) AS VISIT_COUNT
FROM TB_CALL_STORE a WITH(NOLOCK)
cross apply
(
select top 1 prod_id from TB_PRODUCT with(nolock)
)p
inner join TB_TIME d
on a.business_date = d.t_date
where d.section ='2016-11'
GROUP BY customer_id,p.prod_id
)v
group by v.customer_id,
v.prod_id

由于代码本身比较简单，没办法再进一步简化，而由于连接不了服务器，其他的方法也用不上，甚至没办法分析到底是什么导致运行这么慢。

想了想，full join 本质上就是 2次left join+union ，无非就是合并数据，于是尝试一下用union all来直接合并数据，现在改成unoin all最后，就不需要full join。

但是考虑到第2段代码中并没有prod_id这个字段，所以这里在第2段代码加上了cross apply随便取出一个产品的id，这样就有prod_id这个字段，可以合并了。

修改之后，果然速度降到了10多秒。