SQL 2012 SP3 DATEADD 谓词导致计数查询中的高估计

Question

我有最新的 SQL 2012 标准版：

Microsoft SQL Server 2012 (SP3-CU3) (KB3152635) - 11.0.6537.0 (X64) 2016 年 4 月 28 日 17:57:34 版权所有 (c) Microsoft Corporation 标准 Windows NT 6.1（内部版本 7601：Service Pack 1）上的版本（64 位） （管理程序）

根据每个客户的用户数量的 2 个变量（月份和年份）进行简单计数，其中应用程序是特定的，行估计是实际结果的 10 倍。我已经释放了缓存，重建了索引，确保它正在执行索引搜索并尝试重新编译计划。如果我手动输入谓词，我会得到预期的结果和计划，但如果我使用 dateadd，我会得到一个夸张的版本，因为我在使用 CTE 的视图中有多个计数，我正在寻找一种方法来解决大量高估我认为 dateadd 导致的行。

    SELECT
    [Customer],
    [Month],
    [Year],
    COUNT([Username]) AS 'RDS'
    FROM [MS_CitrixUsers]
    WHERE [Application] = 'RDS'
    AND [Month] = DATENAME(month, dateadd(month, -1,GETDATE()))
    AND [YEAR] = DATENAME(year, dateadd(month, -1,GETDATE()))
    GROUP BY [MS_CitrixUsers].[Customer],
    [MS_CitrixUsers].[Month],
    [MS_CitrixUsers].[Year]

计数在包含膨胀行计数的排序之后触发流聚合。

Plan

Stream Properties

我的观点是什么样的：

    WITH RDS1 
    AS
    (
    SELECT
    [Customer],
    COUNT([Username]) AS 'RDS1'
    FROM [MS_CitrixUsers]
    WHERE [Application] = 'RDS'
    AND [Month] = @month 
    AND [YEAR] = @year
    GROUP BY [MS_CitrixUsers].[Customer],
    [MS_CitrixUsers].[Month],
    [MS_CitrixUsers].[Year]
    ),
    RDS2 
    AS
    (
    SELECT
    [Customer],
    [Month],
    [Year],
    COUNT([Username]) AS 'RDS2'
    FROM [MS_CitrixUsers]
    WHERE [Application] = 'RDS'
    AND [Month] = @monthbefore
    AND [YEAR] = @year2
    GROUP BY [MS_CitrixUsers].[Customer],
    [MS_CitrixUsers].[Month],
    [MS_CitrixUsers].[Year]
    ),
    ...

    SELECT c.Customer,
    CASE WHEN NOT EXISTS (SELECT [RDS1] FROM [RDS1] t1 WHERE t1.Customer =     c.Customer) 
AND EXISTS (SELECT [RDS2] FROM [RDS2] t2 WHERE t2.Customer = c.Customer) THEN '99999' ELSE ([RDS1] - Coalesce([RDS2], 0)) END AS 'RDS',
CASE WHEN NOT EXISTS (SELECT [CitrixUsage1] FROM [CitrixUsage] t1 WHERE t1.Customer = c.Customer) 
AND EXISTS (SELECT [CitrixUsage2] FROM [CitrixUsage2] t2 WHERE t2.Customer = c.Customer) THEN '99999' ELSE ([CitrixUsage1] - Coalesce([CitrixUsage2], 0)) END    AS 'Citrix Usage',

我得到的预期结果视图是 2 个月的数据之间的差异：

[客户] [应用程序1] ... [应用程序N]

客户1 -1 3

我们有很多应用程序，因此您可以想象返回结果需要相当长的时间，所以我看看优化估算是否会有所帮助。

解决方法：我发现在这里的另一个答案中，使用临时表，将我的变量粘贴在那里，然后加入我的表以避免流聚合，而是使用哈希匹配来估计正确的行数。

Answer 1

您应该尽量不要在谓词中进行计算。你没有提到这是否是一个视图，但你可以不声明变量吗？

DECLARE @MonthVariable varchar(10); SET @MonthVariable = DATENAME(month, dateadd(month, -1,GETDATE()))
DECLARE @YearVariable int; SET @YearVariable = DATENAME(year, dateadd(month, -1,GETDATE()))

SELECT
    [Customer],
    [Month],
    [Year],
    COUNT([Username]) AS 'RDS'
FROM [MS_CitrixUsers]
WHERE [Application] = 'RDS'
    AND [Month] = @MonthVariable
    AND [YEAR] = @YearVariable
GROUP BY [MS_CitrixUsers].[Customer],
    [MS_CitrixUsers].[Month],
    [MS_CitrixUsers].[Year]

所以这是一个视图，试试这个；

SELECT
    [Customer],
    [Month],
    [Year],
    COUNT([Username]) AS 'RDS'
FROM [MS_CitrixUsers] a
INNER JOIN (SELECT
             DATENAME(month, dateadd(month, -1,GETDATE())) CurMon
            ,DATENAME(year, dateadd(month, -1,GETDATE())) CurYear
            ) b
    ON a.[Month] = b.CurMon
    AND a.[Year] = b.CurYear
WHERE [Application] = 'RDS'
GROUP BY [MS_CitrixUsers].[Customer],
    [MS_CitrixUsers].[Month],
     [MS_CitrixUsers].[Year]

Answer 2

我的解决方法版本：

CREATE TABLE #MasterDate( [Month1] varchar(50), [YEAR1] varchar(10),[Month2] varchar(50),[YEAR2] varchar(10))
INSERT INTO #MasterDate ([Month1],[YEAR1],[Month2],[YEAR2] )
    SELECT DATENAME(month, dateadd(month, -1,GETDATE())) 
    ,DATENAME(year, dateadd(month, -1,GETDATE()))
    ,DATENAME(month, dateadd(month, -2,GETDATE())) 
    ,DATENAME(year, dateadd(month, -2,GETDATE()))

;WITH RDS1 
AS
(
SELECT
[Customer]
,[MS_CitrixUsers].[Month]
,[MS_CitrixUsers].[YEAR]
,COUNT([Username]) AS 'RDS1'
FROM [MS_CitrixUsers]
INNER JOIN #MasterDate m ON [Month] = m.[Month1] and [YEAR] = m.[YEAR1]
WHERE [Application] = 'RDS'
GROUP BY [MS_CitrixUsers].[Customer],
[MS_CitrixUsers].[Month],
[MS_CitrixUsers].[Year]
),

RDS2 
AS
(
SELECT
[Customer],
[Month],
[Year],
COUNT([Username]) AS 'RDS2'
FROM [MS_CitrixUsers]
INNER JOIN #MasterDate m ON [Month] = m.[Month2] and [YEAR] = m.[YEAR2]
WHERE [Application] = 'RDS'
GROUP BY [MS_CitrixUsers].[Customer],
[MS_CitrixUsers].[Month],
[MS_CitrixUsers].[Year]
)
SELECT c.Customer,
CASE WHEN NOT EXISTS (SELECT [RDS1] FROM [RDS1] t1 WHERE t1.Customer = c.Customer) 
AND EXISTS (SELECT [RDS2] FROM [RDS2] t2 WHERE t2.Customer = c.Customer) THEN '99999' ELSE ([RDS1] - Coalesce([RDS2], 0)) END AS 'RDS'
FROM [Customers] c
INNER JOIN #MasterDate m ON [Month] = m.[Month2] and [YEAR] = m.[YEAR2]
LEFT JOIN RDS1 t1 ON t1.Customer = c.Customer
LEFT JOIN RDS2 t2 ON t2.Customer = c.Customer

DROP TABLE #MasterDate

当然，这不能用作视图，只能用作 SP。现在的问题是，它是否解决了我的速度问题，只需再编辑 50 cte。

编辑：完成了我的 SP 版本的视图，内部连接到 #temp 表而不是 dateadd 视图。

查看完成时间：37 秒

SP：11 秒。