1、 问题
最近项目中发现row_number()函数去重写法性能会比group by性能稍微差一些，原语句如下：
bash
SELECT
COUNT(1)
FROM
(
WITH
ORG AS
(
SELECT
*
FROM
(
SELECT
group_code,
group_name,
ROW_NUMBER() OVER(PARTITION BY group_code ORDER BY NULL) RN
FROM
ORG_BASIS
union all
select
'09' as group_code,
group_name ,
ROW_NUMBER() OVER(PARTITION BY group_code ORDER BY NULL) RN
from
ORG_BASIS
where
group_code = '0900'
)
WHERE
RN = 1
)
select
o1.group_name as pName,
o2.group_name as bName ,
t1.BUREAU_CODE
FROM
TRAN_PLAN t1
LEFT JOIN ORG o1
ON
t1.province_code = o1.group_code
LEFT JOIN ORG o2
ON
t1.BUREAU_CODE = o2.group_code
WHERE
t1.PLAN_NATURE = 1
AND t1.POWER_GRID_FLAG = 1
AND t1.SPECIALITY IN (11, 12, 13)
AND t1.PLAN_END_TIME < DATEADD(YEAR, 1, TO_DATE(?, 'yyyy-mm-dd'))
AND t1.PLAN_END_TIME >= TO_DATE(?, 'yyyy-mm-dd')
AND T1.PROVINCE_CODE = ?
AND T1.BUREAU_CODE = ?
AND
(
t1.PLAN_STATE IN (50, 60, 70, 110)
)
)
t;计划：
单次执行性能是正常的，但一定的压力下，热点就在窗口函数，压测情况如下：
这里平均值达5s左右。
row_number() 是一种 SQL 窗口函数，常用于对分组数据进行排序并生成唯一的行号。大白话就是排序去重。
语句中是对group_code分组生成序号取第一个。group_name和group_code是一一对应的，那么对这两列进行分组就可以满足需求。因此可以用group by替换改写。
2、改写
bash
SELECT
COUNT(1)
FROM
(
WITH
ORG AS
(
SELECT
*
FROM
(
SELECT
group_code,
max(group_name) group_name
FROM
ORG_BASIS
group by group_code
union all
select
'09' as group_code,
max(group_name) group_name
from
ORG_BASIS
where
group_code = '0900'
group by group_code
)
)
select
o1.group_name as pName,
o2.group_name as buName ,
t1.bureau_code
FROM
TRAN_PLAN t1
LEFT JOIN ORG o1
ON
t1.province_code = o1.group_code
LEFT JOIN ORG o2
ON
t1.BUREAU_CODE = o2.group_code
WHERE
t1.PLAN_NATURE = 1
AND t1.POWER_GRID_FLAG = 1
AND t1.SPECIALITY IN (11, 12, 13)
AND t1.PLAN_END_TIME < DATEADD(YEAR, 1, TO_DATE(?, 'yyyy-mm-dd'))
AND t1.PLAN_END_TIME >= TO_DATE(?, 'yyyy-mm-dd')
AND T1.PROVINCE_CODE = ?
AND T1.BUREAU_CODE = ?
AND
(
t1.PLAN_STATE IN (50, 60, 70, 110)
)
)
t;计划：
压测情况
性能比较稳定。
3、小结
Row_number()函数会内部进行排序，因此比group by多了排序内存消耗，所以总的来说，能用group by替换row_number()做去重就选择group by。