Question

使用SQL Server 2008。

我有多个位置，每个位置包含多个部门，每个部门包含多个可以具有零到多个扫描的项目。每次扫描都与特定操作有关，该操作可能有也可能没有截止时间。每个Item也属于属于特定Job的特定Package。每个作业包含一个或多个包含一个或多个项目的包。

+=============+                         +=============+
|  Locations  |                         |     Jobs    |
+=============+                         +=============+
      ^                                       ^
      |                                       |
+=============+     +=============+     +=============+
| Departments | <-- |    Items    | --> |   Packages  |
+=============+     +=============+     +=============+
                          ^
                          |
                    +=============+     +=============+
                    |    Scans    | --> | Operations  |
                    +=============+     +=============+

我要做的是获取按位置和扫描日期分组的作业扫描计数。棘手的部分是我只想计算每个项目的第一次扫描日期/时间，其中操作的截止时间不为空。（注意：扫描肯定不会在表中的日期/时间顺序。）

我的查询给我的结果是正确的，但是当作业的项目数量达到75,000左右时，它会非常缓慢。我正在推动一个新服务器 - 我知道我们的硬件缺乏 - 但是我想知道我在查询中是否有什么东西在阻碍它。

从我可以从执行计划中收集到的一点点来看，查询的大部分成本似乎都在子查询中，以找到每个项目的第一个扫描。它对Operations表索引（ID，Cutoff）执行索引扫描（0％），然后执行惰性假脱机（19％）。它对Scans表索引（ItemID，DateTime，OperationID，ID）执行索引搜索（61％）。后续嵌套循环（内部联接）仅为2％，Top运算符为0％。（不是我真的很了解我刚输入的内容，但我想提供尽可能多的信息......）

以下是查询：

SELECT
    Departments.LocationID
    , DATEADD(dd, 0, DATEDIFF(dd, 0, Scans.DateTime))
    , COUNT(Scans.ItemID) AS [COUNT]
FROM
    Items           
    INNER JOIN Scans
        ON Scans.ID = 
    (
        SELECT TOP 1
            Scans.ID 
        FROM
            Scans
        INNER JOIN Operations
            ON Scans.OperationID = Operations.ID
        WHERE
            Operations.Cutoff IS NOT NULL
            AND Scans.ItemID = Items.ID             
        ORDER BY
            Scans.DateTime
    )
    INNER JOIN Operations
        ON Scans.OperationID = Operations.ID
    INNER JOIN Packages
        ON Items.PackageID = Packages.ID
    INNER JOIN Departments
        ON Items.DepartmentID = Departments.ID      
WHERE
    Packages.JobID = @ID        
GROUP BY
    Departments.LocationID 
    , DATEADD(dd, 0, DATEDIFF(dd, 0, Scans.DateTime));

将返回如下结果的样本：

8   2012-06-08 00:00:00.000 11842
21  2012-06-07 00:00:00.000 502
11  2012-06-12 00:00:00.000 1841
15  2012-06-11 00:00:00.000 4314
16  2012-06-09 00:00:00.000 278
23  2012-06-12 00:00:00.000 1345
6   2012-06-06 00:00:00.000 2005
20  2012-06-08 00:00:00.000 352
14  2012-06-07 00:00:00.000 2408
8   2012-06-11 00:00:00.000 290
19  2012-06-10 00:00:00.000 85
20  2012-06-11 00:00:00.000 5484
7   2012-06-10 00:00:00.000 2389
16  2012-06-06 00:00:00.000 6762
18  2012-06-09 00:00:00.000 4473
14  2012-06-10 00:00:00.000 2364
1   2012-06-11 00:00:00.000 1531
22  2012-06-08 00:00:00.000 14534
5   2012-06-10 00:00:00.000 11908
9   2012-06-12 00:00:00.000 47
19  2012-06-07 00:00:00.000 559
7   2012-06-07 00:00:00.000 2576

这是执行计划（不确定自原始帖子后我改变了什么，但成本％略有不同。但瓶颈似乎仍然在同一区域）： Execution Plan

Answer 1

我对于将此标记为答案我有点怀疑，因为我相信我们仍然可以从查询中挤出一点果汁。但这确实使我的测试运行从22秒降低到6秒（在Scans上添加了索引：OperationID包括DateTime和ItemID）：

WITH CTE AS 
(
    SELECT
        Items.ItemID AS ID          
        , Scans.DateTime AS [DateTime]
        , Operations.Cutoff AS Cutoff           
        , ROW_NUMBER() OVER (PARTITION BY Items.ID ORDER BY Scans.DateTime) AS RN
        FROM
            Items
            INNER JOIN Scans            
                ON Items.ID = Scans.ItemID
            INNER JOIN Operations
                ON Scans.OperationID = Operations.ID
            INNER JOIN Packages
                ON Items.PackageID = Packages.ID
        WHERE
            Operations.Cutoff IS NOT NULL
            AND Packages.JobID = @ID                        
)
SELECT
    Departments.LocationID
    , CTE.DateTime
    , COUNT(Items.ID) AS COUNT
FROM
    Items           
    INNER JOIN CTE
        ON Items.ID = CTE.ID
        AND CTE.RN = 1
    INNER JOIN Packages
        ON Items.PackageID = Packages.ID
    INNER JOIN Departments
        ON Items.DepartmentID = Departments.ID      
WHERE
    Packages.JobID = @ID
GROUP BY
    Departments.LocationID 
    , CTE.DateTime

Answer 2

很难肯定地说，但这样的事情可能表现得更好。我用ROW_NUMBER调用替换了您的嵌套查找。原始查询中的问题是嵌套查找 - 它会杀死你。

注意我面前没有SQL，所以我无法测试它，但我认为它在逻辑上是等价的。

SELECT
    Departments.LocationID
    , DATEADD(dd, 0, DATEDIFF(dd, 0, Scans.DateTime))
    , COUNT(Scans.ItemID) AS [COUNT]
FROM
    Items           
    INNER JOIN Scans
        ON Scans.ItemID = Items.ID
    INNER JOIN Operations
        ON Scans.OperationID = Operations.ID
    INNER JOIN Packages
        ON Items.PackageID = Packages.ID
    INNER JOIN Departments
        ON Items.DepartmentID = Departments.ID      
WHERE
    Operations.Cutoff IS NOT NULL
    AND
    Packages.JobID = @ID
    AND
    ROW_NUMBER () OVER (PARTITION BY Items.ID ORDER BY Scans.DateTime) = 1
GROUP BY
    Departments.LocationID 
    , DATEADD(dd, 0, DATEDIFF(dd, 0, Scans.DateTime));

Answer 3

我很好奇 - 请你运行CROSS APPLY版本吗？

SELECT
    Departments.LocationID
    , DATEADD(dd, 0, DATEDIFF(dd, 0, CA_Scans.DateTime))
    , COUNT(CA_Scans.ItemID) AS [COUNT]
FROM
    Items 
    CROSS APPLY
    (
        SELECT TOP 1
            Scans.ID,
            Scans.OperationID,
            Scans.DateTime
        FROM
            Scans
        INNER JOIN Operations
            ON Scans.OperationID = Operations.ID
        WHERE
            Operations.Cutoff IS NOT NULL
            AND Scans.ItemID = Items.ID             
        ORDER BY
            Scans.DateTime
    ) CA_Scans
    INNER JOIN Operations
        ON CA_Scans.OperationID = Operations.ID
    INNER JOIN Packages
        ON Items.PackageID = Packages.ID
    INNER JOIN Departments
        ON Items.DepartmentID = Departments.ID      
WHERE
    Packages.JobID = @ID        
GROUP BY
    Departments.LocationID 
    , DATEADD(dd, 0, DATEDIFF(dd, 0, CA_Scans.DateTime));

具有一对多关系的多表连接

3 个答案: