TSQL - 递归CTE效率低下 - 需要替代方案

Question

这是一个包含示例数据的表:

DECLARE @TestTable TABLE (
    ItemID INT,
    A INT,
    B INT,
    Month INT)

INSERT INTO @TestTable VALUES (1234, 5, 9, 1)
INSERT INTO @TestTable VALUES (1234, 6, 9, 2)
INSERT INTO @TestTable VALUES (4321, 5, 11, 1)
INSERT INTO @TestTable VALUES (4321, 12, 11, 2)
INSERT INTO @TestTable VALUES (1324, 14, 6, 1)
INSERT INTO @TestTable VALUES (1324, 5, 6, 2)
INSERT INTO @TestTable VALUES (1234, 1, 9, 3)
INSERT INTO @TestTable VALUES (1324, 9, 6, 3)

需要注意的是,B列始终与在此计算中仅使用一次相同,但初始计算需要B列.

我试图在第一行中从A中减去B,然后在后续行中从A中减去先前的行差异.有效地,B - A = C在第一C - A行然后在所有后续行上相关的ItemID.

以下是我期待的结果:

ItemID  A   B   C   Month   RowNumber
1234    5   9   4   1       1
1234    6   9   -2  2       2
1234    1   9   -3  3       3
1324    14  6   -8  1       1
1324    5   6   -13 2       2
1324    9   6   -22 3       3
4321    5   11  6   1       1
4321    12  11  -6  2       2

以下是我如何实现这一目标.

;WITH CTE_TestValue AS (
    SELECT 
        Main.ItemID,
        Main.A,
        Main.B,
        Main.Month,
        ROW_NUMBER() OVER (Partition BY Main.ItemID ORDER BY Main.Month) AS RowNumber
    FROM @TestTable AS Main
),
CTE_TestColumnC AS (
    SELECT 
        MainA.ItemID,
        MainA.A,
        MainA.B,
        (MainA.B - MainA.A) AS C,
        MainA.Month,
        MainA.RowNumber
    FROM CTE_TestValue AS MainA
        WHERE MainA.Rownumber = 1

    UNION ALL

    SELECT 
        MainB.ItemID,
        MainB.A,
        MainB.B,
        (Sub.C - MainB.A) AS C,
        MainB.Month,
        MainB.RowNumber
    FROM CTE_TestValue AS MainB
        INNER JOIN CTE_TestColumnC AS Sub
            ON MainB.RowNumber - 1 = Sub.RowNumber
            AND MainB.ItemID = Sub.ItemID
--      CROSS JOIN CTE_TestColumnC AS Sub
--          WHERE Sub.RowNumber + 1 = MainB.RowNumber
--          AND MainB.ItemID = Sub.ItemID 
)
SELECT 
    Main.ItemID,
    Main.A,
    Main.B,
    Main.C,
    Main.Month,
    Main.RowNumber
FROM CTE_TestColumnC AS Main
ORDER BY ItemID, Month, RowNumber

这在小数据样本上工作正常,但我正在处理大约20,000个ItemId,每个重复10次.它按预期完成所有第一行计算,然后计算时间逐渐增加.

正如你所看到的,我已经尝试过a INNER JOIN和a CROSS JOIN.我相信他们拥有与我给出的参数相同的执行计划CROSS JOIN.

有没有更有效/更有效的方法来实现这一目标？

我昨天允许它运行5个小时,看它是否结束了......它没有.

另一个注意事项:当我在测试数据上使用它时,我SELECT没有使用ORDER希望有助于加快速度.ORDER当我正在检查时,这只是为了我的方便.

Answer 1

您的问题是您使用CTE作为递归CTE的来源.对于递归CTE的每次迭代,您的第一个CTE将执行一次.使用您的测试数据意味着CTE_TestValue创建了8次.

将结果放在CTE_TestValue具有聚簇主键的(RowNumber, ItemID)临时表中,并将该临时表用作递归CTE的数据源CTE_TestColumnC.

还要将递归部分中的连接条件更改为ON MainB.RowNumber = Sub.RowNumber + 1.这将使查询能够使用临时表上的索引.

DECLARE @TestTable TABLE (
    ItemID INT,
    A INT,
    B INT,
    Month INT)

INSERT INTO @TestTable VALUES (1234, 5, 9, 1)
INSERT INTO @TestTable VALUES (1234, 6, 9, 2)
INSERT INTO @TestTable VALUES (4321, 5, 11, 1)
INSERT INTO @TestTable VALUES (4321, 12, 11, 2)
INSERT INTO @TestTable VALUES (1324, 14, 6, 1)
INSERT INTO @TestTable VALUES (1324, 5, 6, 2)
INSERT INTO @TestTable VALUES (1234, 1, 9, 3)
INSERT INTO @TestTable VALUES (1324, 9, 6, 3)

CREATE TABLE #TestValue
(
  ItemID INT,
  A INT,
  B INT,
  Month INT,
  RowNumber INT,
  primary key(RowNumber, ItemID)
)

INSERT INTO #TestValue
SELECT 
    Main.ItemID,
    Main.A,
    Main.B,
    Main.Month,
    ROW_NUMBER() OVER (Partition BY Main.ItemID ORDER BY Main.Month) AS RowNumber
FROM @TestTable AS Main


;WITH CTE_TestColumnC AS (
    SELECT 
        MainA.ItemID,
        MainA.A,
        MainA.B,
        (MainA.B - MainA.A) AS C,
        MainA.Month,
        MainA.RowNumber
    FROM #TestValue AS MainA
        WHERE MainA.Rownumber = 1

    UNION ALL

    SELECT 
        MainB.ItemID,
        MainB.A,
        MainB.B,
        (Sub.C - MainB.A) AS C,
        MainB.Month,
        MainB.RowNumber
    FROM #TestValue AS MainB
        INNER JOIN CTE_TestColumnC AS Sub
            ON MainB.RowNumber = Sub.RowNumber + 1
            AND MainB.ItemID = Sub.ItemID
)
SELECT 
    Main.ItemID,
    Main.A,
    Main.B,
    Main.C,
    Main.Month,
    Main.RowNumber
FROM CTE_TestColumnC AS Main
ORDER BY ItemID, Month, RowNumber

DROP TABLE #TestValue

在查询的查询计划中,问题显示在右下角的表扫描中.使用此测试数据,它将执行8次,总共返回64行:

查询使用临时表计划查询: