SQL SERVER 2012 执行计划走嵌套循环导致性能问题的案例

    开发人员遇到一个及其诡异的的SQL性能问题，这段完整SQL语句如下所示：

declare @UserId             INT
declare @PSANo              VARCHAR(200)
declare @ShipMode           VARCHAR(10)
declare @CY_FLAG            VARCHAR(1)
declare @PO                 VARCHAR(20)
declare @BuyerName          VARCHAR(100)
declare @Destination        VARCHAR(1)
declare @FinalDestination   VARCHAR(40)
declare @Factory            VARCHAR(10)
declare @NoticeDateStart    DATETIME
declare @NoticeDateEnd      DATETIME
declare @EELForwarder       VARCHAR(100)
declare @SortExpression     VARCHAR(100)
declare @RowIndex           INT
declare @PageSize           INT
declare @ExistNoticeKey         varchar(200)
DECLARE @NULLDATE DATETIME

SET @NULLDATE=GETDATE()

set @UserId=39
set @PSANo=''
set @ShipMode=''
set @CY_FLAG=''
set @PO=N''
set @BuyerName=N''
set @Destination=N''
set @FinalDestination=N''
set @Factory=''
set @EELForwarder=N''
set @SortExpression=''
set @RowIndex=0
set @PageSize=10
set @ExistNoticeKey=''



    DECLARE @CountSql NVARCHAR(max)
    DECLARE @DataSql NVARCHAR(max)
    declare @next int
    declare @Where_PSANo varchar(400)
    declare @Index_PSANo varchar(40)
    declare @Where_ExcludeNotcekey varchar(400)

    set @Where_PSANo=''

    SET NOCOUNT ON;

    set @next=1
    while @next<=dbo.Get_StrArrayLength(@PSANo,',')
    begin
       set @Index_PSANo = dbo.Get_StrArrayStrOfIndex(@PSANo,',',@next)
       set @Where_PSANo = @Where_PSANo + ' Or notice.PSA_NO LIKE ''%'+@Index_PSANo+'%'''
       set @next=@next+1
    end



    set @Where_ExcludeNotcekey=''
    if @ExistNoticeKey!=''
    begin
        set @Where_ExcludeNotcekey=' or notice.NOTICE_KEY not in('+ @ExistNoticeKey+')';
        --select @Where_ExcludePSANo
        --print 'OK'
    end




SELECT SUM(ISNULL(FactQty,0)) AS FactQty, NOTICE_KEY INTO #TEMP
FROM
(
    SELECT  A.NOTICE_KEY,SUM(ISNULL(A.FactQty,0)) FactQty  FROM IES.InvoiceFourLine A GROUP BY A.NOTICE_KEY
    UNION ALL
    SELECT A.NoticeKey AS NOTICE_KEY,SUM(ISNULL(A.FactQty,0)) FactQty FROM IES.InvoiceThreeByrFwdChargeLine A GROUP BY A.NoticeKey
) T GROUP BY NOTICE_KEY

SELECT COUNT(*)
FROM IES.ExportNotice notice --WITH (INDEX(PK_EXPORTNOTICE))
LEFT  JOIN #TEMP t ON notice.NOTICE_KEY = T.NOTICE_KEY
WHERE
notice.FACTORY_CD IN(SELECT SiteId FROM DCL.SecurityUserSiteMapping WHERE UserId=39)
AND (ISNULL(notice.FACT_EXPORT_QTY,0)-ISNULL(T.FactQty,0))>0
AND (ISNULL(@PSANo,'')=''  Or notice.PSA_NO LIKE '%%')
AND (ISNULL(@ExistNoticeKey,'')='' )
AND (ISNULL(@ShipMode,'')='' OR  notice.SHIP_MODE_CD=@ShipMode)
AND (ISNULL(@CY_FLAG,'')='' OR notice.CY_FLAG=@CY_FLAG)
AND (ISNULL(@PO,'')='' OR notice.BUYER_PO_NO LIKE '%'+@PO+'%')
AND (ISNULL(@BuyerName,'')='' OR notice.NAME LIKE '%'+@BuyerName+'%')
AND (ISNULL(@Destination,'')='' OR notice.SZ=@Destination)
AND (ISNULL(@FinalDestination,'')='' OR notice.FINAL_DESTINATION LIKE '%'+@FinalDestination+'%')
AND (ISNULL(@Factory,'')='' OR notice.FACTORY_CD=@Factory)
AND (ISNULL(@EELForwarder,'')='' OR notice.EEL_FORWARDER=@EELForwarder)
AND (ISNULL(@NoticeDateStart,'2000-01-01')='2000-01-01')
---AND ( ISNULL(@NoticeDateEnd,'1999-01-01')='1999-01-01')


DROP TABLE #TEMP

案例的环境为SQL SERVER 2012 Standard Edition (64-bit)，具体版本号为11.0.5058.0 ，另外表IES.ExportNotice的数据记录为2万多。表IES.InvoiceThreeByrFwdChargeLine的记录数为1万多，表IES.InvoiceFourLine的记录只有区区几十条。临时表 #TEMP的记录为1万多条。

执行上面SQL语句一般一秒以内完成。但是这段SQL如果将最后注释的条件加上（也就是最后注释的语句取消注释）

SELECT COUNT(*)
FROM IES.ExportNotice notice --WITH (INDEX(PK_EXPORTNOTICE))
LEFT  JOIN #TEMP t ON notice.NOTICE_KEY = T.NOTICE_KEY
WHERE
notice.FACTORY_CD IN(SELECT SiteId FROM DCL.SecurityUserSiteMapping WHERE UserId=39)
AND (ISNULL(notice.FACT_EXPORT_QTY,0)-ISNULL(T.FactQty,0))>0
AND (ISNULL(@PSANo,'')=''  Or notice.PSA_NO LIKE '%%')
AND (ISNULL(@ExistNoticeKey,'')='' )
AND (ISNULL(@ShipMode,'')='' OR  notice.SHIP_MODE_CD=@ShipMode)
AND (ISNULL(@CY_FLAG,'')='' OR notice.CY_FLAG=@CY_FLAG)
AND (ISNULL(@PO,'')='' OR notice.BUYER_PO_NO LIKE '%'+@PO+'%')
AND (ISNULL(@BuyerName,'')='' OR notice.NAME LIKE '%'+@BuyerName+'%')
AND (ISNULL(@Destination,'')='' OR notice.SZ=@Destination)
AND (ISNULL(@FinalDestination,'')='' OR notice.FINAL_DESTINATION LIKE '%'+@FinalDestination+'%')
AND (ISNULL(@Factory,'')='' OR notice.FACTORY_CD=@Factory)
AND (ISNULL(@EELForwarder,'')='' OR notice.EEL_FORWARDER=@EELForwarder)
AND (ISNULL(@NoticeDateStart,'2000-01-01')='2000-01-01')
AND ( ISNULL(@NoticeDateEnd,'1999-01-01')='1999-01-01')

然后执行时发现SQL慢得令人发指，非常的不可以思议。 如果按照我们理解，这个条件( ISNULL(@NoticeDateEnd,'1999-01-01')='1999-01-01') 仅仅相当于一个 1=1 或1=0的条件，怎么会有如此大的性能差距呢？ 查看执行计划后，发现加上这样一个条件后，执行计划完全不同了。

我姑且将执行性能较好的SQL的执行计划叫做Plan A，执行性能很差的SQL的执行计划叫做Plan B

Plan A

Plan B

如上所示，Plan B 看似开销都耗费在键查找那一块，但是如果查看具体信息（如下所示），并无特别地方。

于是我使用HINT，强制在表IES.ExportNotice上走索引PK_EXPORTNOTICE，结果发现执行时，执行速度依然慢的令人发指。我觉得执行计划有些问题，Cost可能并不正确。

SELECT COUNT(*)
FROM IES.ExportNotice notice WITH (INDEX(PK_EXPORTNOTICE))
LEFT  JOIN #TEMP t ON notice.NOTICE_KEY = T.NOTICE_KEY
WHERE
notice.FACTORY_CD IN(SELECT SiteId FROM DCL.SecurityUserSiteMapping WHERE UserId=39)
AND (ISNULL(notice.FACT_EXPORT_QTY,0)-ISNULL(T.FactQty,0))>0
AND (ISNULL(@PSANo,'')=''  Or notice.PSA_NO LIKE '%%')
AND (ISNULL(@ExistNoticeKey,'')='' )
AND (ISNULL(@ShipMode,'')='' OR  notice.SHIP_MODE_CD=@ShipMode)
AND (ISNULL(@CY_FLAG,'')='' OR notice.CY_FLAG=@CY_FLAG)
AND (ISNULL(@PO,'')='' OR notice.BUYER_PO_NO LIKE '%'+@PO+'%')
AND (ISNULL(@BuyerName,'')='' OR notice.NAME LIKE '%'+@BuyerName+'%')
AND (ISNULL(@Destination,'')='' OR notice.SZ=@Destination)
AND (ISNULL(@FinalDestination,'')='' OR notice.FINAL_DESTINATION LIKE '%'+@FinalDestination+'%')
AND (ISNULL(@Factory,'')='' OR notice.FACTORY_CD=@Factory)
AND (ISNULL(@EELForwarder,'')='' OR notice.EEL_FORWARDER=@EELForwarder)
AND (ISNULL(@NoticeDateStart,'2000-01-01')='2000-01-01')
AND ( ISNULL(@NoticeDateEnd,'1999-01-01')='1999-01-01')

于是我将怀疑的地方转移到表连接方式，使用Table HINT，强制下面SQL语句走HASH JOIN，结果SQL一秒钟执行完成。

SELECT COUNT(*)
FROM IES.ExportNotice notice
LEFT HASH JOIN #TEMP t ON notice.NOTICE_KEY = T.NOTICE_KEY
WHERE
notice.FACTORY_CD IN(SELECT SiteId FROM DCL.SecurityUserSiteMapping WHERE UserId=39)
AND (ISNULL(notice.FACT_EXPORT_QTY,0)-ISNULL(T.FactQty,0))>0
AND (ISNULL(@PSANo,'')=''  Or notice.PSA_NO LIKE '%%')
AND (ISNULL(@ExistNoticeKey,'')='' )
AND (ISNULL(@ShipMode,'')='' OR  notice.SHIP_MODE_CD=@ShipMode)
AND (ISNULL(@CY_FLAG,'')='' OR notice.CY_FLAG=@CY_FLAG)
AND (ISNULL(@PO,'')='' OR notice.BUYER_PO_NO LIKE '%'+@PO+'%')
AND (ISNULL(@BuyerName,'')='' OR notice.NAME LIKE '%'+@BuyerName+'%')
AND (ISNULL(@Destination,'')='' OR notice.SZ=@Destination)
AND (ISNULL(@FinalDestination,'')='' OR notice.FINAL_DESTINATION LIKE '%'+@FinalDestination+'%')
AND (ISNULL(@Factory,'')='' OR notice.FACTORY_CD=@Factory)
AND (ISNULL(@EELForwarder,'')='' OR notice.EEL_FORWARDER=@EELForwarder)
AND (ISNULL(@NoticeDateStart,'2000-01-01')='2000-01-01')
AND ( ISNULL(@NoticeDateEnd,'1999-01-01')='1999-01-01')

虽然解决了问题，但是我隐隐觉得这应该是SQL SERVER优化器的某些Bug才导致出现这种特殊的情况。而且执行计划的Cost也完全不准确。让人有点匪夷所思。