执行计划与 ROW_NUMBER答案

【问题标题】：Execution plan vs. ROW_NUMBER执行计划与 ROW_NUMBER
【发布时间】：2014-12-13 05:15:38
【问题描述】：

我正在尝试微调使用 ROW_NUMBER 和许多 PARTITION BY 参数的查询，以从数据库中获取唯一记录。查询如下：

SELECT  SDP.SuspectID                       
        ,SDP.RecordID                   
        ,SDP.Field
        ,SDP.FieldEntryDate
        ,SDP.ScreenDate                 
        ,SDP.SuspectDetails              
        ,CONVERT(VARCHAR(10),SDP.PeriodEndDate,120) AS PeriodEndDate        
        ,ROW_NUMBER() OVER (PARTITION BY SDP.recordID, SDP.Field, SDP.ScreenerID, SDP.PeriodEndDate, SDP.PeriodID, SDP.[Source], SDP.SuspectDetails ORDER BY SDP.UploadDate) AS RowNum
        ,SDP.DatabaseAccountCode
        ,SDP.RecordUpdateType
INTO        #Temp
FROM        dbo.SuspectDataPoint
            AS SDP
LEFT JOIN   dbo.Screener
            AS S
            ON S.screenerID = SDP.screenerID                                                            
WHERE       SDP.ScreenerId not in (719)
            AND S.DatabaseName = 'db'
            AND CONVERT(DATE, SDP.FieldEntryDate) > DATEADD(dd,-24,GETDATE())

我尝试手动查找此查询中效率最低的部分是什么，结果发现使用 ROW_NUMBER 时它的运行时间比没有它时长 10 倍以上。

我还尝试检查执行计划以确认该发现，但有一些我不明白的地方。如果我按照上面给出的方式运行查询，执行计划显示最大的成本是插入临时表 (45%)。当我在没有 ROW_NUMBER 行的情况下运行相同的查询时也是这种情况，只是 % 略有不同（61%）。我知道在第一个查询中，ROW_NUMBER 操作发生在SORT，但与其他查询相比，它相当便宜（25%）。然而，正如我所说，没有和使用ROW_NUMBER 行所花费的实际时间非常不同（平均为 1 秒与 13 秒）。有人可以向我解释吗？我是不是看错了计划？

编辑：添加索引视图

编辑：添加架构和索引脚本

/****** Object:  Table [dbo].[Screener]    Script Date: 10/17/2014 5:53:51 PM ******/
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
SET ANSI_PADDING ON
GO
CREATE TABLE [dbo].[Screener](
    [ScreenerID] [int] NOT NULL,
    [ScreenerName] [varchar](255) NOT NULL,
    [ScreenerDescription] [nvarchar](4000) NULL,
    [Script] [text] NULL,
    [HitRate] [numeric](10, 2) NOT NULL DEFAULT ((1)),
    [CreatedOn] [date] NULL,
    [CreatedBy] [varchar](7) NULL,
    [SuspectReason] [nvarchar](4000) NULL,
    [IsExtremeOutlier] [bit] NOT NULL DEFAULT ((0)),
    [DatabaseName] [varchar](20) NOT NULL,
PRIMARY KEY CLUSTERED 
(
    [ScreenerID] 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]

GO
SET ANSI_PADDING OFF
GO
/****** Object:  Table [dbo].[SuspectDataPoint]    Script Date: 10/17/2014 5:53:52 PM ******/
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
SET ANSI_PADDING ON
GO
CREATE TABLE [dbo].[SuspectDataPoint](
    [SuspectID] [int] IDENTITY(1,1) NOT FOR REPLICATION NOT NULL,
    [RecordID] [varchar](30) NOT NULL,
    [Field] [varchar](50) NOT NULL,
    [ScreenerID] [int] NOT NULL,
    [ScreenDate] [datetime] NOT NULL,
    [SuspectDetails] [nvarchar](4000) NULL,
    [PeriodEndDate] [datetime] NULL,
    [FieldEntryDate] [datetime] NOT NULL,
    [OriginalValue] [nvarchar](4000) NULL,
    [Source] [nvarchar](300) NULL,
    [UniqueSystemID] [int] NOT NULL,
    [DatabaseAccountCode] [varchar](50) NULL,
    [RecordUpdateType] [varchar](60) NULL,
    [UploadDate] [datetime] NOT NULL DEFAULT (getdate()),
    [PeriodID] [varchar](20) NULL,
PRIMARY KEY CLUSTERED 
(
    [SuspectID] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY]

GO
SET ANSI_PADDING OFF
GO
SET ANSI_PADDING ON

GO
/****** Object:  Index [in_n_Screener_DatabaseName]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE NONCLUSTERED INDEX [in_n_Screener_DatabaseName] ON [dbo].[Screener]
(
    [DatabaseName] 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]
GO
SET ANSI_PADDING ON

GO
/****** Object:  Index [in_n_SuspectDataPoint_DatabaseAccountCode]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE NONCLUSTERED INDEX [in_n_SuspectDataPoint_DatabaseAccountCode] ON [dbo].[SuspectDataPoint]
(
    [DatabaseAccountCode] 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]
GO
SET ANSI_PADDING ON

GO
/****** Object:  Index [in_n_SuspectDataPoint_Field]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE NONCLUSTERED INDEX [in_n_SuspectDataPoint_Field] ON [dbo].[SuspectDataPoint]
(
    [Field] 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]
GO
SET ANSI_PADDING ON

GO
/****** Object:  Index [in_n_SuspectDataPoint_RecordID]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE NONCLUSTERED INDEX [in_n_SuspectDataPoint_RecordID] ON [dbo].[SuspectDataPoint]
(
    [RecordID] 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]
GO
/****** Object:  Index [in_n_SuspectDataPoint_ScreenerID]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE NONCLUSTERED INDEX [in_n_SuspectDataPoint_ScreenerID] ON [dbo].[SuspectDataPoint]
(
    [ScreenerID] ASC
)
INCLUDE (   [ScreenDate]) 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]
GO
/****** Object:  Index [in_n_SuspectDataPoint_UniqueSystemID]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE NONCLUSTERED INDEX [in_n_SuspectDataPoint_UniqueSystemID] ON [dbo].[SuspectDataPoint]
(
    [UniqueSystemID] 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]
GO
/****** Object:  Index [SuspectDataPoint_FieldEntryDate_Index]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE NONCLUSTERED INDEX [SuspectDataPoint_FieldEntryDate_Index] ON [dbo].[SuspectDataPoint]
(
    [FieldEntryDate] 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]
GO
SET ANSI_PADDING ON

GO
/****** Object:  Index [SuspectDataPoint_RecordUpdateType_Index]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE NONCLUSTERED INDEX [SuspectDataPoint_RecordUpdateType_Index] ON [dbo].[SuspectDataPoint]
(
    [RecordUpdateType] 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]
GO
/****** Object:  Index [SuspectDataPoint_ScreenDate_Index]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE NONCLUSTERED INDEX [SuspectDataPoint_ScreenDate_Index] ON [dbo].[SuspectDataPoint]
(
    [ScreenDate] 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]
GO
SET ANSI_PADDING ON

GO
/****** Object:  Index [SuspectDataPoint_Source_Index]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE NONCLUSTERED INDEX [SuspectDataPoint_Source_Index] ON [dbo].[SuspectDataPoint]
(
    [Source] 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]
GO
/****** Object:  Index [SuspectDataPoint_SuspectID_Index]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE UNIQUE NONCLUSTERED INDEX [SuspectDataPoint_SuspectID_Index] ON [dbo].[SuspectDataPoint]
(
    [SuspectID] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, IGNORE_DUP_KEY = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
GO
/****** Object:  Index [SuspectDataPoint_UploadDate_Index]    Script Date: 10/17/2014 5:53:52 PM ******/
CREATE NONCLUSTERED INDEX [SuspectDataPoint_UploadDate_Index] ON [dbo].[SuspectDataPoint]
(
    [UploadDate] 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]
GO
ALTER TABLE [dbo].[Screener]  WITH CHECK ADD  CONSTRAINT [FK_Screener_DatabaseInfo] FOREIGN KEY([DatabaseName])
REFERENCES [dbo].[DatabaseInfo] ([DatabaseName])
ON UPDATE CASCADE
GO
ALTER TABLE [dbo].[Screener] CHECK CONSTRAINT [FK_Screener_DatabaseInfo]
GO

【问题讨论】：

您能否将您的架构和索引设置添加到问题中？
糟糕，抱歉，编辑不正确。我在查询中没有看到这两个数据库的任何架构，我会尽快发布它们的索引
虽然图像有助于快速概览，但最好为表和索引添加 SQL 脚本。通过右键单击 db 并选择 tasks/generate scripts 并在以下选项卡的高级对话框中选择表并添加索引，可以轻松生成代码。您可以输出到剪贴板并粘贴。
不确定您是否收到有关我的编辑的通知；添加了请求的信息。
我确实收到了通知，添加的信息有所帮助。一会儿再看

标签： sql-server tsql sql-execution-plan

【解决方案1】：

我怀疑并行度会扭曲执行计划中的百分比。

另一方面，“计算标量”发生在“with ROW_NUMBER”计划中的并行排序之后，这可能意味着

CONVERT(DATE, SDP.FieldEntryDate) > DATEADD(dd,-24,GETDATE())

where 子句的一部分发生在 Row_Number OVER 子句之后。如果那个 WHERE 子句排除了 SPD 表的很大一部分，那就可以解释为什么 OVER 子句对查询的影响如此之大。您可以先（拥有）过滤器（ed），然后运行 ROW_NUMBER。

【讨论】：