今天同事咨询一个SQL语句,如下所示,SQL语句本身并不复杂,但是执行效率非常糟糕,糟糕到一塌糊涂(执行计划也是相当复杂)。如果查询条件中没有NOT EXISTS部分,倒是不要一秒就能查询出来。
SELECT * FROM dbo.UVW_PDATest a WITH(NOLOCK)
WHERE
Remark='前纺' AND Operation_Name='粗纱' AND One_Status_Code='0047'
AND a.Createtime >='2015-9-23'
AND NOT EXISTS
(
SELECT 1 FROM dbo.UVW_PDATest c WITH(NOLOCK)
WHERE a.Task_NO =c.Task_NO AND c.One_Status_Code='0014'
)
为什么如此简单的SQL语句,执行效率却一塌糊涂呢,因为UVW_PDATest是一个视图,而且该视图是由8个表关联组成。
SELECT ..........
From dbo.PDA_TB_Produce a With(Nolock)
Join dbo.DctOperationList b With(Nolock)
On a.Operation_Code=b.Operation_Code
Join dbo.DctOneStatusList c With(Nolock)
On a.One_Status_Code=c.One_Status_Code
Left join dbo.DctTwoStatusList d With(Nolock)
On c.One_Status_Code=d.One_Status_Code and a.Two_Status_Code=d.Two_Status_Code
left Join dbo.DctMachineList e With(Nolock)
On a.Operation_Code=e.Operation_Code and a.Machine_Code=e.Machine_Code
left Join dbo.DctOperationList f With(Nolock)
On a.Next_Operation_Code=f.Operation_Code
Join dbo.DctUserList g With(Nolock)
On a.User_ID_Operating=g.User_ID
Join dbo.DctUserList h With(Nolock)
On a.User_ID=h.User_ID
刚开始我想从索引上去优化,加上一两个索引后发现其实并无多大益处。为什么性能会如此糟糕呢?原因是什么呢?
大量复杂的Join
该类查询模式包含了大量连接,尤其是连接条件是不等连接,由于统计信息随着表连接的增多精度逐渐下降,这会导致低效的查询性能。解决这类情况可以通过分解查询,并将中间解决存入临时表解决。 具体参考博客:什么情况下应该分解复杂的查询来提升性能
于是我拆分上面SQL语句(如下所示),先将执行结果保存到临时表,然后关联取数,结果一秒钟的样子就执行出来了。真可谓是化繁为简。
SELECT Task_NO INTO #TMP_MID_UVW_PDATest
FROM dbo.UVW_PDATest c WITH(NOLOCK)
WHERE One_Status_Code='0014' and Remark='前纺' AND Operation_Name='粗纱'
SELECT * INTO #TMP_UVW_PDATest
FROM dbo.UVW_PDATest a WITH(NOLOCK)
WHERE Remark='前纺'
AND Operation_Name='粗纱'
AND One_Status_Code='0047'
AND Create_Date>='2015-9-23' ;
SELECT * FROM #TMP_UVW_PDATest a
WHERE NOT EXISTS(SELECT 1 FROM #TMP_MID_UVW_PDATest c WHERE a.Task_NO =c.Task_NO );
DROPTABLE#TMP_UVW_PDATest
DROP TABLE #TMP_MID_UVW_PDATest
第二个案例是ORACLE数据库的一个优化案例,具体SQL语句如下所示,执行时间非常长,一般都是二十多秒左右。
SELECT A.SC_NO,
A.MRP_GROUP_CD,
A.DIMM_ID,
A.JOB_ORDER_NO,
DECODE(SIGN(A.DEMAND_QTY),-1,0,A.DEMAND_QTY) AS DIFF_QTY,
A.ASSIGNED_TYPE
FROM
(
SELECT CC.SC_NO,
BB.MRP_GROUP_CD,
BB.DIMM_ID,
BB.JOB_ORDER_NO,
NVL (SUM (BB.DEMAND_QTY), 0) - NVL(SUM(REC.RECV_QTY),0) AS DEMAND_QTY,
CASE
WHEN DD.REQ_DATE<TRUNC(SYSDATE) THEN 'AH'
ELSE 'AS'
END AS ASSIGNED_TYPE
FROM MRP_JO_DEMAND BB,
PO_HD CC ,
(
SELECT JOB_ORDER_NO,
DIMM_ID,
SUM(RECV_QTY) AS RECV_QTY
FROM MRP_AGPO_SCHD_RECV_SPECIFIC
GROUP BY JOB_ORDER_NO,
DIMM_ID
)
REC,
MRP_JO_ASSIGN DD
WHERE BB.JOB_ORDER_NO=CC.PO_NO
AND BB.JOB_ORDER_NO=REC.JOB_ORDER_NO(+)
AND BB.DIMM_ID=REC.DIMM_ID(+)
AND BB.JOB_ORDER_NO = DD.JOB_ORDER_NO(+)
AND BB.DIMM_ID = DD.DIMM_ID(+)
AND BB.MRP_GROUP_CD=DD.MRP_GROUP_CD(+)
AND EXISTS
(
SELECT 1
FROM MRP_DIMM AA
WHERE AA.MRP_GROUP_CD=BB.MRP_GROUP_CD
AND AA.DIMM_ID=BB.DIMM_ID
AND AA.JOB_ORDER_NO=BB.JOB_ORDER_NO
)
GROUP BY CC.SC_NO,
BB.MRP_GROUP_CD,
BB.DIMM_ID,
BB.JOB_ORDER_NO,
DD.REQ_DATE
)
A,
INVSUBMAT.INV_MRP_JO_AVAILABLE_V B
WHERE A.JOB_ORDER_NO = B.JOB_ORDER_NO
AND A.MRP_GROUP_CD = B.MRP_GROUP_CD
AND A.DIMM_ID = B.DIMM_ID
AND NVL (A.DEMAND_QTY, 0) < NVL (B.AVAILABLE_QTY, 0)
AND NVL (B.AVAILABLE_QTY, 0)>0
ORDER BY A.MRP_GROUP_CD,
A.DIMM_ID,
A.JOB_ORDER_NO;
查看执行计划,你会发现COST主要耗费在HASH JOIN上。如下截图所示,表INV_STOCK_ASSIGN来自于视图INVSUBMAT.INV_MRP_JO_AVAILABLE_V。
将上面复杂SQL拆分后,执行只需要不到一秒解决,如下截图所示,速率提高了几十倍。优化往往有时候很复杂,有时候也很简单,就是将复杂的语句拆分成简单的SQL语句,性能的提升有时候确实令人吃惊!
CREATE GLOBAL TEMPORARY TABLE TMP_MRP_MID_DATA
( SC_NO VARCHAR2(20) ,
MRP_GROUP_CD VARCHAR2(10) ,
DIMM_ID NUMBER,
JOB_ORDER_NO VARCHAR2(20) ,
DEMAND_QTY NUMBER ,
DIFF_QTY NUMBER ,
ASSIGNED_TYPE VARCHAR(2)
) ON COMMIT PRESERVE ROWS;
INSERT INTO TMP_MRP_MID_DATA
SELECT A.SC_NO,
A.MRP_GROUP_CD,
A.DIMM_ID,
A.JOB_ORDER_NO,
A.DEMAND_QTY,
DECODE(SIGN(A.DEMAND_QTY),-1,0,A.DEMAND_QTY) AS DIFF_QTY,
A.ASSIGNED_TYPE
FROM
(
SELECT CC.SC_NO,
BB.MRP_GROUP_CD,
BB.DIMM_ID,
BB.JOB_ORDER_NO,
NVL (SUM (BB.DEMAND_QTY), 0) - NVL(SUM(REC.RECV_QTY),0) AS DEMAND_QTY,
CASE
WHEN DD.REQ_DATE<TRUNC(SYSDATE) THEN 'AH'
ELSE 'AS'
END AS ASSIGNED_TYPE
FROM MRP_JO_DEMAND BB
INNER JOIN PO_HD CC ON BB.JOB_ORDER_NO=CC.PO_NO
LEFT JOIN (
SELECT JOB_ORDER_NO,
DIMM_ID,
SUM(RECV_QTY) AS RECV_QTY
FROM MRP_AGPO_SCHD_RECV_SPECIFIC
GROUP BY JOB_ORDER_NO,
DIMM_ID
)
REC ON BB.JOB_ORDER_NO=REC.JOB_ORDER_NO AND BB.DIMM_ID=REC.DIMM_ID
LEFT JOIN MRP_JO_ASSIGN DD ON BB.JOB_ORDER_NO = DD.JOB_ORDER_NO AND BB.DIMM_ID = DD.DIMM_ID AND BB.MRP_GROUP_CD=DD.MRP_GROUP_CD
INNER JOIN MRP_DIMM AA ON AA.MRP_GROUP_CD=BB.MRP_GROUP_CD AND AA.DIMM_ID=BB.DIMM_ID AND AA.JOB_ORDER_NO=BB.JOB_ORDER_NO
GROUP BY CC.SC_NO,
BB.MRP_GROUP_CD,
BB.DIMM_ID,
BB.JOB_ORDER_NO,
DD.REQ_DATE
)
A;
COMMIT;
SELECT A.* FROM
TMP_MRP_MID_DATA A INNER JOIN
INVSUBMAT.INV_MRP_JO_AVAILABLE_V B ON A.JOB_ORDER_NO = B.JOB_ORDER_NO
AND A.MRP_GROUP_CD = B.MRP_GROUP_CD
AND A.DIMM_ID = B.DIMM_ID
WHERE
NVL (A.DEMAND_QTY, 0) < NVL (B.AVAILABLE_QTY, 0)
AND NVL (B.AVAILABLE_QTY, 0)>0
ORDER BY A.MRP_GROUP_CD,
A.DIMM_ID,
A.JOB_ORDER_NO;
小结:
1:越是复杂的SQL语句,优化器越是容易选择一个糟糕的执行计划(优化器之所以难以选定最优的执行计划,是因为优化器要平衡选定最优执行路径的代价,不能一味为了选择最优执行计划,而将复杂SQL的所有执行路径都计算对比一遍,往往只能有选择性的选取一些执行路径计算对比,否则开销太大。而越是复杂的SQL,可选择的执行路径就是越多。
说得有点绕口,还是打个比方,比如你从广州到北京,如果就只有飞机(直飞),火车(直达)、汽车(直达)三种选择,那么想必你能很快给出一个最优的路线(例如,最快的是飞机、最省钱的是火车),但是如果飞机、火车、汽车都不能直达:假如火车票没有了直达,你必须中途转几次、飞机票也没有直达了,你需要转机,那么此时选择性复杂的情况,你就必须花费不少时间才能制定一个最优的计划了。 如果在复杂一点的情况,你从中国去美国,是不是有N种路径? 如果全部计算对比一遍各种可能的路径,估计你小脑袋不够用………………
2:执行计划是可以被重用的,越简单的SQL语句被重用的可能性越高。而复杂的SQL语句只要有一个字符发生变化就必须重新解析,然后再把这一大堆垃圾塞在内存里。可想而知,数据库的效率会何等低下。
3:如果SQL语句过分复杂,要么是业务有问题,要么是模型设计不当。可以说复杂的SQL反映出系统设计方面有不少问题和缺陷。