从视图中选择 * 需要 4 分钟

Question

我遇到了一个问题，当我对视图运行查询时，它需要 4 分钟以上的时间。但是，当我运行查询的内容时，它会在 1 秒内完成。

我唯一不确定的是被连接的表都是时态表。

临时查询计划：https : //www.brentozar.com/pastetheplan/?id=BykohB2p4

查看查询计划：https : //www.brentozar.com/pastetheplan/?id=SkIfTHh6E

关于在哪里尝试解决这个问题的任何建议？

查看代码：

ALTER VIEW [dbo].[vwDealHistoryPITA]
AS
SELECT ROW_NUMBER() OVER (PARTITION BY cm.CodeMasterID ORDER BY cm.CodeMasterID, cm.LastUpdateDate) AS Deal_HistoryID,
       cm.CodeMasterID,
       cm.ProjectName,
       cm.[Status],
       d.CompanyID,
       d.DealTypeMasterID,
       cm.[Description],
       d.PassiveInd,
       d.ApproxTPGOwnership,
       d.NumberBoardSeats,
       d.FollowonInvestmentInd,
       d.SocialImpactInd,
       d.EquityInd,
       d.DebtInd,
       d.RealEstateInd,
       d.TargetPctgReturn,
       d.ApproxTotalDealSize,
       cm.CurrencyCode,
       d.ConflictCheck,
       cm.CreatedDate,
       cm.CreatedBy,
       cm.LastUpdateDate,
       cm.LastUpdateBy,
       d.ExpensesExceedThresholdDate,
       d.CurrentTPGCheckSize,
       d.PreferredEquityInd,
       d.ConvertibleDebtInd,
       d.OtherRealAssetsInd,
       d.InitialTPGCheckSize,
       d.DirectLendingInd,
       cm.NameApproved,
       cm.FolderID,
       cm.CodaProcessedDateTime,
       cm.DeadDate,
       d.SectorMasterID,
       d.DTODataCompleteDate,
       cm.ValidFrom AS CodeMasterValidFrom,
       cm.ValidTo   AS CodeMasterValidTo,
       d.validFrom  AS DealValidFrom,
       d.validTo    AS DealValidTo
FROM dbo.CodeMaster FOR SYSTEM_TIME ALL cm 
INNER JOIN dbo.Deal FOR SYSTEM_TIME ALL d ON cm.CodeMasterID = d.CodeMasterID;
GO

添加 Partition by 并获得与临时查询相似的结果。

Answer 1

主要性能差异

这里的主要区别在于，性能更好的查询是CodeMasterID在所有 4 个表（2 个临时表（实际和历史））上下推搜索谓词，其中视图上的选择似乎直到结束（过滤器运算符）才这样做。

特尔博士；

问题是由于在某些情况下（例如视图）参数没有下推到窗口函数。最简单的解决方案是添加OPTION(RECOMPILE)到视图调用中，以便优化器在运行时“看到”参数（如果可能的话）。如果为每个查询调用重新编译执行计划的成本太高，使用需要参数的内联表值函数可能是一种解决方案。有一位优秀的博文由保罗·怀特这一点。有关查找和解决特定问题的更详细方法，请继续阅读。

性能更好的查询

代码管理员表

交易表

我喜欢早上搜索谓词的味道

大坏查询

代码管理员表

这是一个仅限谓词的区域

交易表

但是优化器没有读取“交易的艺术™”

...并且不吸取过去的教训

直到所有数据到达过滤器运算符

那么，什么给？

这里的主要问题是优化器由于视图中的窗口函数而无法在运行时“看到”参数，并且无法使用SelOnSeqPrj （选择序列项目，在这篇文章中进一步介绍以供参考）。

我能够使用测试样本复制相同的结果并使用SP_EXECUTESQL参数化对视图的调用。请参阅 DDL / DML 的附录

使用窗口函数和一个对测试视图执行查询 INNER JOIN

SET STATISTICS IO, TIME ON;
EXEC SP_EXECUTESQL
N'SELECT * FROM  dbo.Bad
Where CodeMasterID = @P1',N'@P1 INT',@P1 = 37155;

导致大约 4.5 秒的 CPU 时间和 3.2 秒的经过时间

 SQL Server Execution Times:
   CPU time = 4595 ms,  elapsed time = 3209 ms.

当我们加入甜蜜的拥抱 OPTION(RECOMPILE)

SET STATISTICS IO, TIME ON;
EXEC SP_EXECUTESQL
N'SELECT * FROM  dbo.Bad
Where CodeMasterID = @P1 OPTION(RECOMPILE)',N'@P1 INT',@P1 = 37155; 

这一切都很好。

 SQL Server Execution Times:
   CPU time = 16 ms,  elapsed time = 98 ms.

为什么

这一切再次支持无法将@P1谓词应用于表的观点，因为窗口函数和参数化导致过滤器运算符

不仅是临时表的问题

见附录 2

即使不使用时态表，也会发生这种情况：

编写这样的查询时会看到相同的结果：

DECLARE @P1 int = 37155
SELECT * FROM  dbo.Bad2
Where CodeMasterID = @P1;

同样，优化器在应用窗口函数之前不会下推谓词。

省略 ROW_NUMBER() 时

CREATE VIEW dbo.Bad3
as
SELECT
cm.CodeMasterID,CM.ManagerID,cm.ParentDeptID,d.DealID, d.CodeMasterID as dealcodemaster,d.EvenMoreBlaID
FROM dbo.CodeMaster2  cm 
INNER JOIN dbo.Deal2  d ON cm.CodeMasterID = d.CodeMasterID;

一切都很好

SET STATISTICS IO, TIME ON;
EXEC SP_EXECUTESQL
N'SELECT * FROM  dbo.Bad3
Where CodeMasterID = @P1',N'@P1 INT',@P1 = 37155


 SQL Server Execution Times:
   CPU time = 0 ms,  elapsed time = 33 ms.

那么这一切让我们何去何从呢？

本ROW_NUMBER()就不好了查询应用过滤器之前的计算方法。

所有这些都让我们看到了Paul White 于2013 年撰写的关于窗口函数和视图的博文。

我们示例的重要部分之一是以下语句：

不幸的是，SelOnSeqPrj 简化规则仅在谓词与常量进行比较时才有效。因此，以下查询在 SQL Server 2008 及更高版本上生成次优计划：

DECLARE @ProductID INT = 878;

SELECT
    mrt.ProductID,
    mrt.TransactionID,
    mrt.ReferenceOrderID,
    mrt.TransactionDate,
    mrt.Quantity
FROM dbo.MostRecentTransactionsPerProduct AS mrt 
WHERE
    mrt.ProductID = @ProductID;

这部分对应于我们在自己声明参数/SP_EXECUTESQL在视图上使用时所看到的。

实际解决方案

1：选项（重新编译）

我们知道OPTION(RECOMPILE)在运行时“看到”值是可能的。当为每个查询调用重新编译执行计划成本太高时，还有其他解决方案。

2：带参数的内联表值函数

CREATE FUNCTION dbo.BlaBla
(
    @P1 INT
)  
RETURNS TABLE
WITH SCHEMABINDING AS
RETURN
    (
     SELECT 
     ROW_NUMBER() OVER (PARTITION BY cm.CodeMasterID ORDER BY cm.CodeMasterID) AS Deal_HistoryID,
     cm.CodeMasterID,CM.ManagerID,
     cm.ParentDeptID,d.DealID,
     d.CodeMasterID as dealcodemaster,
     d.EvenMoreBlaID
    FROM dbo.CodeMaster2  cm 
    INNER JOIN dbo.Deal2  d ON cm.CodeMasterID = d.CodeMasterID
    Where cm.CodeMasterID = @P1
    ) 
EXEC SP_EXECUTESQL
N'SELECT * FROM  dbo.BlaBLa(@P1)',N'@P1 INT',@P1 = 37155

导致预期的搜索谓词

     SQL Server Execution Times:
       CPU time = 0 ms,  elapsed time = 0 ms.

在我的测试中大约有 9 个逻辑读取

3：在不使用视图的情况下编写查询。

另一个“解决方案”可能是完全不使用视图来编写查询。

4：不在ROW_NUMBER()视图中保留函数，而是在调用视图时指定它。

这方面的一个例子是：

CREATE VIEW dbo.Bad2
as
SELECT 
cm.CodeMasterID,CM.ManagerID,cm.ParentDeptID,d.DealID, d.CodeMasterID as dealcodemaster,d.EvenMoreBlaID
FROM dbo.CodeMaster2  cm 
INNER JOIN dbo.Deal2  d ON cm.CodeMasterID = d.CodeMasterID;

GO
SET STATISTICS IO, TIME ON;
EXEC SP_EXECUTESQL
N'SELECT ROW_NUMBER() OVER (PARTITION BY CodeMasterID ORDER BY CodeMasterID) AS Deal_HistoryID,* FROM  dbo.Bad2
Where CodeMasterID = @P1',N'@P1 INT',@P1 = 37155;

应该有其他创造性的方法来解决这个问题，重要的部分是知道是什么原因造成的。

附录#1

CREATE TABLE dbo.Codemaster   
(    
     CodeMasterID int NOT NULL PRIMARY KEY CLUSTERED  
   , ManagerID INT  NULL  
   , ParentDeptID int NULL  
   , SysStartTime datetime2 GENERATED ALWAYS AS ROW START NOT NULL  
   , SysEndTime datetime2 GENERATED ALWAYS AS ROW END NOT NULL  
   , PERIOD FOR SYSTEM_TIME (SysStartTime,SysEndTime)     
)    
WITH (SYSTEM_VERSIONING = ON (HISTORY_TABLE = dbo.Codemaster_History))   
;  

CREATE TABLE dbo.Deal   
(    
     DealID int NOT NULL PRIMARY KEY CLUSTERED  
   , CodeMasterID INT  NULL  
   , EvenMoreBlaID int NULL  
   , SysStartTime datetime2 GENERATED ALWAYS AS ROW START NOT NULL  
   , SysEndTime datetime2 GENERATED ALWAYS AS ROW END NOT NULL  
   , PERIOD FOR SYSTEM_TIME (SysStartTime,SysEndTime)     
)    
WITH (SYSTEM_VERSIONING = ON (HISTORY_TABLE = dbo.Deal_History))   
;  

INSERT INTO dbo.Codemaster(CodeMasterID,ManagerID,ParentDeptID)
SELECT TOP(1000000) ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum1,
                    ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum2,
                    ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum3
FROM MASTER..spt_values as spt1
CROSS JOIN MASTER..spt_values as spt2;


INSERT INTO dbo.Deal(DealID,CodeMasterID,EvenMoreBlaID)
SELECT TOP(1000000) ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum1,
                    ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum2,
                    ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum3
FROM MASTER..spt_values as spt1
CROSS JOIN MASTER..spt_values as spt2;

CREATE INDEX IX_CodeMasterID
ON dbo.Deal(CodeMasterId);
CREATE INDEX IX_CodeMasterID
ON dbo.Deal_History(CodeMasterId);

CREATE INDEX IX_CodeMasterID
ON dbo.Codemaster(CodeMasterId);
CREATE INDEX IX_CodeMasterID
ON dbo.Codemaster_History(CodeMasterId);


SELECT ROW_NUMBER() OVER (PARTITION BY cm.CodeMasterID ORDER BY cm.CodeMasterID, cm.SysStartTime) AS Deal_HistoryID,
cm.*, d.* 
FROM dbo.CodeMaster FOR SYSTEM_TIME ALL cm 
INNER JOIN dbo.Deal FOR SYSTEM_TIME ALL d ON cm.CodeMasterID = d.CodeMasterID
Where cm.CodeMasterID = 37155;

-- Guud
GO
CREATE VIEW dbo.Bad
as
SELECT ROW_NUMBER() OVER (PARTITION BY cm.CodeMasterID ORDER BY cm.CodeMasterID, cm.SysStartTime) AS Deal_HistoryID,
cm.CodeMasterID,CM.ManagerID,cm.ParentDeptID,d.DealID, d.CodeMasterID as dealcodemaster,d.EvenMoreBlaID
FROM dbo.CodeMaster FOR SYSTEM_TIME ALL cm 
INNER JOIN dbo.Deal FOR SYSTEM_TIME ALL d ON cm.CodeMasterID = d.CodeMasterID

GO
EXEC SP_EXECUTESQL
N'SELECT * FROM  dbo.Bad
Where CodeMasterID = @P1',N'@P1 INT',@P1 = 37155

-- Very bad shame on you

附录#2

CREATE TABLE dbo.Codemaster2
(    
     CodeMasterID int NOT NULL PRIMARY KEY CLUSTERED  
   , ManagerID INT  NULL  
   , ParentDeptID int NULL  

);  

CREATE TABLE dbo.Deal2
(    
     DealID int NOT NULL PRIMARY KEY CLUSTERED  
   , CodeMasterID INT  NULL  
   , EvenMoreBlaID int NULL    
);  

INSERT INTO dbo.Codemaster2(CodeMasterID,ManagerID,ParentDeptID)
SELECT TOP(1000000) ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum1,
                    ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum2,
                    ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum3
FROM MASTER..spt_values as spt1
CROSS JOIN MASTER..spt_values as spt2;


INSERT INTO dbo.Deal2(DealID,CodeMasterID,EvenMoreBlaID)
SELECT TOP(1000000) ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum1,
                    ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum2,
                    ROW_NUMBER() OVER(ORDER BY(SELECT NULL)) as rownum3
FROM MASTER..spt_values as spt1
CROSS JOIN MASTER..spt_values as spt2;

CREATE INDEX IX_CodeMasterID
ON dbo.Deal2(CodeMasterId);
CREATE INDEX IX_CodeMasterID
ON dbo.Codemaster2(CodeMasterId);


SELECT ROW_NUMBER() OVER (PARTITION BY cm.CodeMasterID ORDER BY cm.CodeMasterId) AS Deal_HistoryID,
cm.*, d.* 
FROM dbo.CodeMaster2 cm 
INNER JOIN dbo.Deal2 d ON cm.CodeMasterID = d.CodeMasterID
Where cm.CodeMasterID = 37155;

-- Guud
GO
CREATE VIEW dbo.Bad2
as
SELECT ROW_NUMBER() OVER (PARTITION BY cm.CodeMasterID ORDER BY cm.CodeMasterID) AS Deal_HistoryID,
cm.CodeMasterID,CM.ManagerID,cm.ParentDeptID,d.DealID, d.CodeMasterID as dealcodemaster,d.EvenMoreBlaID
FROM dbo.CodeMaster2  cm 
INNER JOIN dbo.Deal2  d ON cm.CodeMasterID = d.CodeMasterID

GO
SET STATISTICS IO, TIME ON;
EXEC SP_EXECUTESQL
N'SELECT * FROM  dbo.Bad2
Where CodeMasterID = @P1',N'@P1 INT',@P1 = 37155