非常相似的查询，截然不同的性能

Question

非常相似的查询，截然不同的性能

Cho*_*man 9 performance sql-server optimization execution-plan query-performance

我有两个非常相似的查询

第一个查询：

SELECT count(*)
FROM Audits a
    JOIN AuditRelatedIds ari ON a.Id = ari.AuditId
WHERE 
    ari.RelatedId = '1DD87CF1-286B-409A-8C60-3FFEC394FDB1'
    and a.TargetTypeId IN 
    (1,2,3,4,5,6,7,8,9,
    11,12,13,14,15,16,17,18,19,
    21,22,23,24,25,26,27,28,29,30,
    31,32,33,34,35,36,37,38,39,
    41,42,43,44,45,46,47,48,49,
    51,52,53,54,55,56,57,58,59,
    61,62,63,64,65,66,67,68,69,
    71,72,73,74,75,76,77,78,79)

Run Code Online (Sandbox Code Playgroud)

结果：267479

计划：https : //www.brentozar.com/pastetheplan/?id=BJWTtILyS

第二个查询：

SELECT count(*)
FROM Audits a
    JOIN AuditRelatedIds ari ON a.Id = ari.AuditId
WHERE 
    ari.RelatedId = '1DD87CF1-286B-409A-8C60-3FFEC394FDB1'
    and a.TargetTypeId IN 
    (1,2,3,4,5,6,7,8,9,
    11,12,13,14,15,16,17,18,19,
    21,22,23,24,25,26,27,28,29,
    31,32,33,34,35,36,37,38,39,
    41,42,43,44,45,46,47,48,49,
    51,52,53,54,55,56,57,58,59,
    61,62,63,64,65,66,67,68,69,
    71,72,73,74,75,76,77,78,79)

Run Code Online (Sandbox Code Playgroud)

结果：25650

计划：https : //www.brentozar.com/pastetheplan/?id=S1v79U8kS

第一个查询大约需要一秒钟才能完成，而第二个查询大约需要 20 秒。这对我来说完全违反直觉，因为第一个查询的计数比第二个查询高得多。这是在 SQL Server 2012 上

为什么差别这么大？如何将第二个查询加速到与第一个查询一样快？

这是两个表的创建表脚本：

CREATE TABLE [dbo].[AuditRelatedIds](
    [AuditId] [bigint] NOT NULL,
    [RelatedId] [uniqueidentifier] NOT NULL,
    [AuditTargetTypeId] [smallint] NOT NULL,
 CONSTRAINT [PK_AuditRelatedIds] PRIMARY KEY CLUSTERED 
(
    [AuditId] ASC,
    [RelatedId] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY]

CREATE NONCLUSTERED INDEX [IX_AuditRelatedIdsRelatedId_INCLUDES] ON [dbo].[AuditRelatedIds]
(
    [RelatedId] ASC
)
INCLUDE (   [AuditId]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]

ALTER TABLE [dbo].[AuditRelatedIds]  WITH CHECK ADD  CONSTRAINT [FK_AuditRelatedIds_AuditId_Audits_Id] FOREIGN KEY([AuditId])
REFERENCES [dbo].[Audits] ([Id])

ALTER TABLE [dbo].[AuditRelatedIds] CHECK CONSTRAINT [FK_AuditRelatedIds_AuditId_Audits_Id]

ALTER TABLE [dbo].[AuditRelatedIds]  WITH CHECK ADD  CONSTRAINT [FK_AuditRelatedIds_AuditTargetTypeId_AuditTargetTypes_Id] FOREIGN KEY([AuditTargetTypeId])
REFERENCES [dbo].[AuditTargetTypes] ([Id])

ALTER TABLE [dbo].[AuditRelatedIds] CHECK CONSTRAINT [FK_AuditRelatedIds_AuditTargetTypeId_AuditTargetTypes_Id]

Run Code Online (Sandbox Code Playgroud)

CREATE TABLE [dbo].[Audits](
    [Id] [bigint] IDENTITY(1,1) NOT NULL,
    [TargetTypeId] [smallint] NOT NULL,
    [TargetId] [nvarchar](40) NOT NULL,
    [TargetName] [nvarchar](max) NOT NULL,
    [Action] [tinyint] NOT NULL,
    [ActionOverride] [tinyint] NULL,
    [Date] [datetime] NOT NULL,
    [UserDisplayName] [nvarchar](max) NOT NULL,
    [DescriptionData] [nvarchar](max) NULL,
    [IsNotification] [bit] NOT NULL,
 CONSTRAINT [PK_Audits] PRIMARY KEY CLUSTERED 
(
    [Id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY] TEXTIMAGE_ON [PRIMARY]

SET ANSI_PADDING ON

CREATE NONCLUSTERED INDEX [IX_AuditsTargetId] ON [dbo].[Audits]
(
    [TargetId] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]

SET ANSI_PADDING ON

CREATE NONCLUSTERED INDEX [IX_AuditsTargetTypeIdAction_INCLUDES] ON [dbo].[Audits]
(
    [TargetTypeId] ASC,
    [Action] ASC
)
INCLUDE (   [TargetId],
    [UserDisplayName]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON, FILLFACTOR = 100) ON [PRIMARY]

ALTER TABLE [dbo].[Audits]  WITH CHECK ADD  CONSTRAINT [FK_Audits_TargetTypeId_AuditTargetTypes_Id] FOREIGN KEY([TargetTypeId])
REFERENCES [dbo].[AuditTargetTypes] ([Id])

ALTER TABLE [dbo].[Audits] CHECK CONSTRAINT [FK_Audits_TargetTypeId_AuditTargetTypes_Id]

Run Code Online (Sandbox Code Playgroud)

Answer 1

Ran*_*gen 8

Tl;博士在底部

为什么选择了糟糕的计划

选择一种计划而不是另一种计划的主要原因是Estimated total subtree成本。

与性能较好的计划相比，糟糕计划的成本更低。

坏计划的总估计子树成本：

更好执行计划的总估计子树成本

运营商预估成本

某些运营商可以承担大部分成本，这可能是优化器选择不同路径/计划的原因。

在我们更好的执行计划中，大部分Subtreecost是根据index seek&nested loops operator执行连接计算的：

而对于我们糟糕的查询计划，Clustered index seek算子成本更低

这应该解释为什么可以选择另一个计划。

（并通过添加参数来30增加已超出871.510000估计成本的坏计划的成本）。 估计猜测™

更好的执行计划

糟糕的计划

这会把我们带到哪里？

此信息使我们能够在我们的示例中强制执行错误的查询计划 （有关用于复制问题的数据，请参阅 DML 以几乎复制 OP 的问题）

通过添加INNER LOOP JOIN连接提示

SELECT count(*)
FROM Audits a
   INNER LOOP JOIN AuditRelatedIds ari ON a.Id = ari.AuditId
WHERE 
    ari.RelatedId = '1DD87CF1-286B-409A-8C60-3FFEC394FDB1'
    and a.TargetTypeId IN 
    (1,2,3,4,5,6,7,8,9,
    11,12,13,14,15,16,17,18,19,
    21,22,23,24,25,26,27,28,29,
    31,32,33,34,35,36,37,38,39,
    41,42,43,44,45,46,47,48,49,
    51,52,53,54,55,56,57,58,59,
    61,62,63,64,65,66,67,68,69,
    71,72,73,74,75,76,77,78,79)

Run Code Online (Sandbox Code Playgroud)

它更接近，但有一些连接顺序差异：

重写

我的第一次重写尝试可能是将所有这些数字存储在一个临时表中：

CREATE TABLE #Numbers(Numbering INT)
INSERT INTO #Numbers(Numbering)
VALUES
(1),(2),(3),(4),(5),(6),(7),(8),(9),(11),(12),(13),(14),(15),(16),(17),(18),(19),
(21),(22),(23),(24),(25),(26),(27),(28),(29),(30),(31),(32),(33),(34),(35),
(36),(37),(38),(39),(41),(42),(43),(44),(45),(46),(47),(48),(49),(51),(52),
(53),(54),(55),(56),(57),(58),(59),(61),(62),(63),(64),(65),(66),(67),(68),
(69),(71),(72),(73),(74),(75),(76),(77),(78),(79);

Run Code Online (Sandbox Code Playgroud)

然后添加一个JOIN而不是大IN()

SELECT count(*)
FROM Audits a
   INNER LOOP JOIN AuditRelatedIds ari ON a.Id = ari.AuditId
   INNER JOIN #Numbers
   ON Numbering = a.TargetTypeId
WHERE 
    ari.RelatedId = '1DD87CF1-286B-409A-8C60-3FFEC394FDB1';

Run Code Online (Sandbox Code Playgroud)

我们的查询计划不同但尚未修复：

AuditRelatedIds桌上有巨大的估计运营商成本

这是我注意到的地方

我不能直接重新创建您的计划的原因是优化的位图过滤。

我可以通过使用 traceflags 7497&禁用优化的位图过滤器来重新创建您的计划7498

SELECT count(*)
FROM Audits a 
   INNER JOIN AuditRelatedIds  ari ON a.Id = ari.AuditId 
   INNER JOIN #Numbers
   ON Numbering = a.TargetTypeId
WHERE 
    ari.RelatedId = '1DD87CF1-286B-409A-8C60-3FFEC394FDB1'
OPTION (QUERYTRACEON 7497, QUERYTRACEON 7498);

Run Code Online (Sandbox Code Playgroud)

有关优化位图过滤器的更多信息，请访问此处。

这意味着，如果没有位图过滤器，优化器认为最好先连接到#number表，然后再连接到AuditRelatedIds表。

在强制执行命令时， OPTION (QUERYTRACEON 7497, QUERYTRACEON 7498, FORCE ORDER);我们可以看到原因：

&

不好

删除与 maxdop 1 并行的能力

添加MAXDOP 1查询时执行速度更快，单线程。

并添加此索引

CREATE NONCLUSTERED INDEX [IX_AuditRelatedIdsRelatedId_AuditId] ON [dbo].[AuditRelatedIds]
(
    [RelatedId] ASC,
    [AuditId] ASC
) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY];

Run Code Online (Sandbox Code Playgroud)

使用合并连接时。

当我们删除强制顺序查询提示或不使用#Numbers 表并使用IN()代替时，情况也是如此。

我的建议是考虑添加MAXDOP(1)并查看是否对您的查询有帮助，并在需要时进行重写。

当然，您还应该记住，在我看来，由于优化了位图过滤并且实际上使用了多线程以达到良好的效果，它的性能甚至更好：

TL; 博士

估计成本将定义所选择的计划，我能够复制行为并看到optimized bitmap filters+parallellism运算符添加到我的最后以高性能和快速的方式执行查询。

您可以考虑添加MAXDOP(1)到您的查询中，希望每次都能获得相同的受控结果，merge join并且没有 'bad' parallellism。

升级到更新的版本并使用更高的基数估计器版本 CardinalityEstimationModelVersion="70"也可能有所帮助。

用于多值过滤的数字临时表也有帮助。

DML 几乎复制了 OP 的问题

我花在这上面的时间比我愿意承认的要多

set NOCOUNT ON;
DECLARE @I INT = 0
WHILE @I < 56
BEGIN
INSERT INTO  [dbo].[Audits] WITH(TABLOCK) 
([TargetTypeId],
    [TargetId],
    [TargetName],
    [Action],
    [ActionOverride] ,
    [Date] ,
    [UserDisplayName],
    [DescriptionData],
    [IsNotification]) 
SELECT top(500000) CASE WHEN ROW_NUMBER() OVER(ORDER BY (SELECT NULL)) / 10000 = 30 then 29 ELSE ROW_NUMBER() OVER(ORDER BY (SELECT NULL)) / 10000 END as rownum2 -- TILL 50 and no 30
,'bla','bla2',1,1,getdate(),'bla3','Bla4',1
FROM master.dbo.spt_values spt1
CROSS APPLY master.dbo.spt_values spt2;
SET @I +=1;
END

-- 'Bad Query matches'
INSERT INTO  [dbo].[AuditRelatedIds] WITH(TABLOCK)
    ([AuditId] ,
    [RelatedId]  ,
    [AuditTargetTypeId])
SELECT
TOP(25650)
ROW_NUMBER() OVER(ORDER BY (SELECT NULL)) as rownum1, 
('1DD87CF1-286B-409A-8C60-3FFEC394FDB1') , 
CASE WHEN ROW_NUMBER() OVER(ORDER BY (SELECT NULL)) / 510 = 30 then 29 ELSE ROW_NUMBER() OVER(ORDER BY (SELECT NULL)) / 510 END as rownum2 -- TILL 50 and no 30
FROM master.dbo.spt_values spt1
CROSS APPLY master.dbo.spt_values spt2

-- Extra matches with 30
SELECT MAX([Id]) FROM [dbo].[Audits];
--28000001 Upper value

INSERT INTO  [dbo].[Audits] WITH(TABLOCK) 
([TargetTypeId],
    [TargetId],
    [TargetName],
    [Action],
    [ActionOverride] ,
    [Date] ,
    [UserDisplayName],
    [DescriptionData],
    [IsNotification]) 
SELECT top(241829) 30 as rownum2 -- TILL 50 and no 30
,'bla','bla2',1,1,getdate(),'bla3','Bla4',1
FROM master.dbo.spt_values spt1
CROSS APPLY master.dbo.spt_values spt2;



;WITH CTE AS
(SELECT
ROW_NUMBER() OVER(ORDER BY (SELECT NULL)) as rownum1, 
('1DD87CF1-286B-409A-8C60-3FFEC394FDB1') as gu , 
30 as rownum2 -- TILL 50 and no 30
FROM master.dbo.spt_values spt1
CROSS APPLY master.dbo.spt_values spt2
CROSS APPLY master.dbo.spt_values spt3
)
--267479 - 25650 = 241829
INSERT INTO  [dbo].[AuditRelatedIds] WITH(TABLOCK)
    ([AuditId] ,
    [RelatedId]  ,
    [AuditTargetTypeId])

SELECT TOP(241829) rownum1,gu,rownum2 FROM CTE
WHERE rownum1 > 28000001
ORDER BY rownum1 ASC;

Run Code Online (Sandbox Code Playgroud)

归档时间：	6 年，4 月前
查看次数：	236 次
最近记录：	6 年，4 月前