T-SQL 流程设计和执行计划（UDF 参数嗅探？）
Posted
技术标签:
【中文标题】T-SQL 流程设计和执行计划（UDF 参数嗅探？）【英文标题】：T-SQL Process Design and Execution Plan (UDF Parameter Sniffing?) 【发布时间】：2009-05-15 02:08:53 【问题描述】：在 SQL Server 2005 上，我有一个复杂的多级分配过程，看起来像这样（伪 SQL）：
FOR EACH @LVL_NUM < @MAX_LVL:
INSERT INTO ALLOCS
SELECT 'OUT', *
FROM BALANCES(@LVL_NUM)
INNER JOIN ALLOCN_SUMRY(@LVL_NUM)
INSERT INTO ALLOCS
SELECT 'IN', *
FROM BALANCES(@LVL_NUM)
INNER JOIN ALLOCNS(@LVL_NUM)
INNER JOIN ALLOCN_SUMRY(@LVL_NUM)
其中ALLOCS 是直接分配种子，然后BALANCES(@LVL_NUM) 基于@LVL_NUM 上的ALLOCS（可能是一些直接分配加上前一级的一些IN 分配），ALOCNS(@LVL_NUM) 是基于在BALANCES(@LVL_NUM) 和ALOCN_SUMRY(@LVL_NUM) 上只是基于ALOCNS(@LVL_NUM) - 有很多配置表指示驱动分配的驱动程序。
这个是简化的，但实际上循环中有四五对这样的对，因为有多种逻辑不能一起处理（有些情况可以一起处理）。
基本逻辑是在特定成本中心/产品线/等（即BALANCES）中获取总金额，然后根据其份额将其分配到另一个成本中心/产品线/等（即ALLOCNS / ALLOCN_SUMRY百分比份额）特定指标。
在OUT 记录保存和IN 中重复了如此多的逻辑，当然还有基于ALLOCN 细节的SUMRY，我最终使用内联表值函数来实现，这似乎执行得相当好（并且它们在回归测试中匹配现有系统的行为，这是一个加号！）。 （现有的系统是一个怪物C/C++/MFC/ODBC程序，将所有数据读入海量数组和其他数据结构中，写得非常糟糕。）
问题似乎是，当在循环中运行时，我似乎遇到了执行计划问题，因为ALLOCS 表开始发生变化（一切都在变化，因为级别不同）成本中心，因此用于驱动 ALLOCNS 的配置正在改变）。我认为我最多有 99 个级别，但最低级别从 2、4、6 开始。似乎在 UDF 之外单独运行 @LVL_NUM = 6 性能很好，但 UDF 性能很差 - 大概是因为 UDF 有一个缓存的计划或由于 @LVL_NUM IN (2, 4) 前面步骤中添加的 ALLOCS 导致整体计划已经很糟糕。
在开发早期，我设法在 30 分钟内完成了 30 个关卡，但现在我无法在 2 小时内完成前 3 个关卡。
我正在考虑在另一个 SP 中运行这两个插入并将其称为 WITH RECOMPILE，但很好奇这个 RECOMPILE 是否正确级联到 TVF UDF 中？任何其他建议也将不胜感激。
真实代码：
/****** Object: UserDefinedFunction [MISProcess].[udf_MR_BALANCES_STAT_UNI] Script Date: 05/14/2009 22:16:09 ******/
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
CREATE FUNCTION [MISProcess].[udf_MR_BALANCES_STAT_UNI] (
@DATA_DT_ID int
,@LVL_NUM int
)
RETURNS TABLE
-- WITH SCHEMABINDING
AS
RETURN
(
SELECT AB.YYMM_ID
,AB.BUS_UNIT_ID
,AB.BUS_UNIT_PROD_LINE_CD
-- ,AB.ALOCN_SRC_CD
,AB.ALOCN_SRC_PROD_LINE_CD
,CASE WHEN ORIG_ALSRC.ALOCN_TYPE_CD = 'C'
AND ORIG_ALSRC.RETN_IND = 'Y' THEN AB.ORIG_ALOCN_SRC_CD
ELSE AB.BUS_UNIT_ID
END AS ORIG_ALOCN_SRC_CD
,CASE WHEN BUPALSRC.COLLAPSE_IND = 'Y'
THEN BUPLNTM.ALOCN_LINE_ITEM_NUM
ELSE AB.LINE_ITEM_NUM
END AS ALOCN_LINE_ITEM_NUM
,SUM(BUPLNTM.ALOCN_SIGN_IND * AB.ANULZD_ACTL_BAL) AS ANULZD_ACTL_BAL
FROM MISWork.vwMR_BALANCES AS AB
INNER JOIN MISProcess.LKP_BUPLNTM AS BUPLNTM
ON BUPLNTM.DATA_DT_ID = @DATA_DT_ID
AND BUPLNTM.LINE_ITEM_NUM = AB.LINE_ITEM_NUM
AND BUPLNTM.ALOCN_LINE_ITEM_NUM <> 0
INNER JOIN [MISProcess].[udf_MR_ALSRC](@DATA_DT_ID, @LVL_NUM) AS BUPALSRC
ON BUPALSRC.ALOCN_SRC_CD = AB.BUS_UNIT_ID
INNER JOIN [MISProcess].LKP_BUPALSRC AS ORIG_ALSRC
ON ORIG_ALSRC.DATA_DT_ID = @DATA_DT_ID
AND ORIG_ALSRC.ALOCN_SRC_CD = AB.ORIG_ALOCN_SRC_CD
GROUP BY AB.YYMM_ID
,AB.BUS_UNIT_ID
,AB.BUS_UNIT_PROD_LINE_CD
-- ,AB.ALOCN_SRC_CD
,AB.ALOCN_SRC_PROD_LINE_CD
,CASE WHEN ORIG_ALSRC.ALOCN_TYPE_CD = 'C'
AND ORIG_ALSRC.RETN_IND = 'Y' THEN AB.ORIG_ALOCN_SRC_CD
ELSE AB.BUS_UNIT_ID
END
,CASE WHEN BUPALSRC.COLLAPSE_IND = 'Y'
THEN BUPLNTM.ALOCN_LINE_ITEM_NUM
ELSE AB.LINE_ITEM_NUM
END
)
/****** Object: UserDefinedFunction [MISProcess].[udf_MR_ALOCNS_STAT_UNI] Script Date: 05/14/2009 22:16:16 ******/
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
CREATE FUNCTION [MISProcess].[udf_MR_ALOCNS_STAT_UNI] (
@DATA_DT_ID int
,@LVL_NUM int
)
RETURNS TABLE
-- WITH SCHEMABINDING
AS
RETURN
(
SELECT BALANCES.YYMM_ID
,BS.ALOCN_SRC_CD AS BUS_UNIT_ID
,BS.PROD_LINE_CD AS BUS_UNIT_PROD_LINE_CD
,BALANCES.BUS_UNIT_ID AS ALOCN_SRC_CD
,BALANCES.BUS_UNIT_PROD_LINE_CD AS ALOCN_SRC_PROD_LINE_CD
,BALANCES.ORIG_ALOCN_SRC_CD
,BALANCES.ALOCN_LINE_ITEM_NUM
,SUM(BS.ACCT_STATS_CNT) AS ACCT_STATS_CNT
FROM [MISProcess].[udf_MR_BALANCES_STAT_UNI](@DATA_DT_ID, @LVL_NUM) AS BALANCES
INNER JOIN [MISProcess].[udf_MR_ALSRC](@DATA_DT_ID, @LVL_NUM) AS BUPALSRC
ON BUPALSRC.ALOCN_SRC_CD = BALANCES.BUS_UNIT_ID
INNER JOIN MISProcess.LKP_PRODLINE AS PRODLINE
ON PRODLINE.DATA_DT_ID = @DATA_DT_ID
AND PRODLINE.PROD_LINE_CD = BALANCES.BUS_UNIT_PROD_LINE_CD
INNER JOIN PUASFIN.FocusResults.BS AS BS
ON BS.YYMM_ID = BALANCES.YYMM_ID
AND BS.ALOCN_BASE_CD = BUPALSRC.ALOCN_BASE_CD
AND BS.ALOCN_SRC_CD <> BALANCES.BUS_UNIT_ID
AND (
PRODLINE.GENRC_PROD_LINE_IND = 'Y'
OR BS.PROD_LINE_CD = BALANCES.BUS_UNIT_PROD_LINE_CD
)
INNER JOIN [MISProcess].[udf_MR_ALSRC](@DATA_DT_ID, 0) AS DEST_BUP_ALSRC
ON DEST_BUP_ALSRC.ALOCN_SRC_CD = BS.ALOCN_SRC_CD
AND DEST_BUP_ALSRC.ALOCN_LVL_NUM > @LVL_NUM
LEFT JOIN [MISProcess].[udf_MR_BLOCK_STD_COST_PCT](@DATA_DT_ID) AS BLOCK_STD_COST_PCT
ON BLOCK_STD_COST_PCT.FROM_ALOCN_SRC_CD = BALANCES.BUS_UNIT_ID
LEFT JOIN [MISProcess].[udf_MR_BLOCK_NOT](@DATA_DT_ID) AS BLOCK_NOT
ON BLOCK_NOT.ALOCN_SRC_CD = BALANCES.BUS_UNIT_ID
LEFT JOIN [MISProcess].[udf_MR_BLOCK](@DATA_DT_ID) AS BLOCK
ON BLOCK_NOT.ALOCN_SRC_CD IS NULL
AND BLOCK.FROM_ALOCN_SRC_CD = BALANCES.BUS_UNIT_ID
AND (
BLOCK.FROM_PROD_LINE_CD IS NULL
OR BLOCK.FROM_PROD_LINE_CD = BALANCES.BUS_UNIT_PROD_LINE_CD
)
LEFT JOIN [MISProcess].[udf_MR_BLOCK_ALOCN_PAIRS](@DATA_DT_ID, @LVL_NUM)
AS BLOCK_ALOCN_PAIRS
ON BLOCK_NOT.ALOCN_SRC_CD IS NOT NULL
AND BLOCK_ALOCN_PAIRS.FROM_ALOCN_SRC_CD = BALANCES.BUS_UNIT_ID
AND BLOCK_ALOCN_PAIRS.TO_ALOCN_SRC_CD = BS.ALOCN_SRC_CD
WHERE BLOCK_ALOCN_PAIRS.TO_ALOCN_SRC_CD IS NULL
AND BLOCK_STD_COST_PCT.FROM_ALOCN_SRC_CD IS NULL
AND (
BLOCK.TO_ALOCN_SRC_CD IS NULL
OR BLOCK.TO_ALOCN_SRC_CD = BS.ALOCN_SRC_CD
)
AND (
BLOCK.TO_PROD_LINE_CD IS NULL
OR BLOCK.TO_PROD_LINE_CD = BS.PROD_LINE_CD
)
AND (
BLOCK.YEAR_NUM IS NULL
OR BLOCK.YEAR_NUM = BALANCES.YYMM_ID / 10000
)
AND (
BLOCK.MTH_NUM IS NULL
OR BLOCK.MTH_NUM = (BALANCES.YYMM_ID / 100) % 100
)
AND (
BLOCK.TO_DIV_NUM IS NULL
OR BLOCK.TO_DIV_NUM = DEST_BUP_ALSRC.DIV_NUM
)
AND (
BLOCK.TO_GRP_NUM IS NULL
OR BLOCK.TO_GRP_NUM = DEST_BUP_ALSRC.DIV_GRP
)
AND (
BLOCK.TO_REGN_GRP_NM IS NULL
OR BLOCK.TO_REGN_GRP_NM = DEST_BUP_ALSRC.REGN_GRP_NM
)
AND (
BLOCK.TO_REGN_NM IS NULL
OR BLOCK.TO_REGN_NM = DEST_BUP_ALSRC.REGN_NM
)
AND (
BLOCK.TO_ARENA_NM IS NULL
OR BLOCK.TO_ARENA_NM = DEST_BUP_ALSRC.ARENA_NM
)
AND (
BLOCK.TO_SUB_REGN_NM IS NULL
OR BLOCK.TO_SUB_REGN_NM = DEST_BUP_ALSRC.SUB_REGN_NM
)
AND (
BLOCK.TO_SUB_ARENA_NM IS NULL
OR BLOCK.TO_SUB_ARENA_NM = DEST_BUP_ALSRC.SUB_ARENA_NM
)
GROUP BY BALANCES.YYMM_ID
,BS.ALOCN_SRC_CD
,BS.PROD_LINE_CD
,BALANCES.BUS_UNIT_ID
,BALANCES.BUS_UNIT_PROD_LINE_CD
,BALANCES.ORIG_ALOCN_SRC_CD
,BALANCES.ALOCN_LINE_ITEM_NUM
)
/****** Object: UserDefinedFunction [MISProcess].[udf_MR_ALOCN_SUMRY_STAT_UNI] Script Date: 05/14/2009 22:16:28 ******/
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
CREATE FUNCTION [MISProcess].[udf_MR_ALOCN_SUMRY_STAT_UNI] (
@DATA_DT_ID int
,@LVL_NUM int
)
RETURNS TABLE
-- WITH SCHEMABINDING
AS
RETURN
(
SELECT YYMM_ID
,ALOCN_SRC_CD
,ALOCN_SRC_PROD_LINE_CD
,ORIG_ALOCN_SRC_CD
,ALOCN_LINE_ITEM_NUM
,SUM(ACCT_STATS_CNT) AS ACCT_STATS_CNT
FROM [MISProcess].[udf_MR_ALOCNS_STAT_UNI](@DATA_DT_ID, @LVL_NUM) AS ALOCNS
GROUP BY YYMM_ID
,ALOCN_SRC_CD
,ALOCN_SRC_PROD_LINE_CD
,ORIG_ALOCN_SRC_CD
,ALOCN_LINE_ITEM_NUM
)
这是我的测试批次，最终将在单个 SP 中运行整个过程。您可以从注释掉的部分看到我也一直在使用临时表和表变量：
USE PCAPFIN
DECLARE @DATA_DT_ID_use AS int
DECLARE @MinLevel AS int
DECLARE @MaxLevel AS int
DECLARE @TestEveryLevel AS bit
DECLARE @TestFinal AS bit
SET @DATA_DT_ID_use = 20090331
SET @MinLevel = 6
SET @MaxLevel = 6
SET @TestEveryLevel = 0
SET @TestFinal = 1
--DECLARE @BALANCES TABLE (
-- METHOD_TXT varchar(12) NOT NULL
-- ,YYMM_ID int NOT NULL
-- ,BUS_UNIT_ID varchar(6) NOT NULL
-- ,BUS_UNIT_PROD_LINE_CD varchar(4) NOT NULL
-- ,ALOCN_SRC_PROD_LINE_CD varchar(4) NOT NULL
-- ,ORIG_ALOCN_SRC_CD varchar(6) NOT NULL
-- ,ALOCN_LINE_ITEM_NUM int NOT NULL
-- ,ANULZD_ACTL_BAL money
-- )
--
--DECLARE @ALOCNS TABLE (
-- METHOD_TXT varchar(12) NOT NULL
-- ,YYMM_ID int NOT NULL
-- ,BUS_UNIT_ID varchar(6) NOT NULL
-- ,BUS_UNIT_PROD_LINE_CD varchar(4) NOT NULL
-- ,ALOCN_SRC_CD varchar(6) NOT NULL
-- ,ALOCN_SRC_PROD_LINE_CD varchar(4) NOT NULL
-- ,ORIG_ALOCN_SRC_CD varchar(6) NOT NULL
-- ,ALOCN_LINE_ITEM_NUM int NOT NULL
-- ,ACCT_STATS_CNT money
-- )
--
--DECLARE @ALOCN_SUMRY TABLE (
-- METHOD_TXT varchar(12) NOT NULL
-- ,YYMM_ID int NOT NULL
-- ,ALOCN_SRC_CD varchar(6) NOT NULL
-- ,ALOCN_SRC_PROD_LINE_CD varchar(4) NOT NULL
-- ,ORIG_ALOCN_SRC_CD varchar(6) NOT NULL
-- ,ALOCN_LINE_ITEM_NUM int NOT NULL
-- ,ACCT_STATS_CNT money
-- )
--IF OBJECT_ID('tempdb..#BALANCES') IS NOT NULL
-- DROP TABLE #BALANCES
--
--CREATE TABLE #BALANCES (
-- METHOD_TXT varchar(12) NOT NULL
-- ,YYMM_ID int NOT NULL
-- ,BUS_UNIT_ID varchar(6) NOT NULL
-- ,BUS_UNIT_PROD_LINE_CD varchar(4) NOT NULL
-- ,ALOCN_SRC_PROD_LINE_CD varchar(4) NOT NULL
-- ,ORIG_ALOCN_SRC_CD varchar(6) NOT NULL
-- ,ALOCN_LINE_ITEM_NUM int NOT NULL
-- ,ANULZD_ACTL_BAL money
-- ,CONSTRAINT [PK_BALANCES] PRIMARY KEY CLUSTERED ([METHOD_TXT] ASC, [YYMM_ID] ASC, [BUS_UNIT_ID] ASC, [BUS_UNIT_PROD_LINE_CD] ASC, [ALOCN_SRC_PROD_LINE_CD] ASC, [ORIG_ALOCN_SRC_CD] ASC, [ALOCN_LINE_ITEM_NUM] ASC)
-- WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF,
-- IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON,
-- ALLOW_PAGE_LOCKS = ON)
-- )
--
--IF OBJECT_ID('tempdb..#ALOCN_SUMRY') IS NOT NULL
-- DROP TABLE #ALOCNS
--
--CREATE TABLE #ALOCNS (
-- METHOD_TXT varchar(12) NOT NULL
-- ,YYMM_ID int NOT NULL
-- ,BUS_UNIT_ID varchar(6) NOT NULL
-- ,BUS_UNIT_PROD_LINE_CD varchar(4) NOT NULL
-- ,ALOCN_SRC_CD varchar(6) NOT NULL
-- ,ALOCN_SRC_PROD_LINE_CD varchar(4) NOT NULL
-- ,ORIG_ALOCN_SRC_CD varchar(6) NOT NULL
-- ,ALOCN_LINE_ITEM_NUM int NOT NULL
-- ,ACCT_STATS_CNT money
-- ,CONSTRAINT [PK_ALOCNS] PRIMARY KEY CLUSTERED ([METHOD_TXT] ASC, YYMM_ID ASC, BUS_UNIT_ID ASC, BUS_UNIT_PROD_LINE_CD ASC, ALOCN_SRC_CD ASC, ALOCN_SRC_PROD_LINE_CD ASC, ORIG_ALOCN_SRC_CD ASC, ALOCN_LINE_ITEM_NUM ASC)
-- WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF,
-- IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON,
-- ALLOW_PAGE_LOCKS = ON)
-- )
--
--IF OBJECT_ID('tempdb..#ALOCN_SUMRY') IS NOT NULL
-- DROP TABLE #ALOCN_SUMRY
--CREATE TABLE #ALOCN_SUMRY (
-- METHOD_TXT varchar(12) NOT NULL
-- ,YYMM_ID int NOT NULL
-- ,ALOCN_SRC_CD varchar(6) NOT NULL
-- ,ALOCN_SRC_PROD_LINE_CD varchar(4) NOT NULL
-- ,ORIG_ALOCN_SRC_CD varchar(6) NOT NULL
-- ,ALOCN_LINE_ITEM_NUM int NOT NULL
-- ,ACCT_STATS_CNT money
-- ,CONSTRAINT [PK_ALOCN_SUMRY] PRIMARY KEY CLUSTERED ([METHOD_TXT] ASC, YYMM_ID ASC, ALOCN_SRC_CD ASC, ALOCN_SRC_PROD_LINE_CD ASC, ORIG_ALOCN_SRC_CD ASC, ALOCN_LINE_ITEM_NUM ASC)
-- WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF,
-- IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON,
-- ALLOW_PAGE_LOCKS = ON)
-- )
SET @MinLevel = (
SELECT MIN(BUPALSRC.ALOCN_LVL_NUM)
FROM MISProcess.LKP_BUPALSRC AS BUPALSRC
WHERE BUPALSRC.DATA_DT_ID = @DATA_DT_ID_use
AND BUPALSRC.ALOCN_LVL_NUM >= @MinLevel
)
DECLARE @Restart AS bit
IF @MinLevel > (
SELECT MIN(BUPALSRC.ALOCN_LVL_NUM)
FROM MISProcess.LKP_BUPALSRC AS BUPALSRC
WHERE BUPALSRC.DATA_DT_ID = @DATA_DT_ID_use
)
SET @Restart = 0
ELSE
SET @Restart = 1
DECLARE @subset_criteria AS varchar(max)
SET NOCOUNT ON
IF @Restart = 1
BEGIN
RAISERROR ('Restarting process', 10, 1) WITH NOWAIT
-- TRUNCATE TABLE MISWork.AB
DELETE FROM MISWork.AB
INSERT INTO MISWork.AB (
YYMM_ID
,BUS_UNIT_ID
,BUS_UNIT_PROD_LINE_CD
,ALOCN_SRC_CD
,ALOCN_SRC_PROD_LINE_CD
,ORIG_ALOCN_SRC_CD
,LINE_ITEM_NUM
,BAL_ORIGTN_IND
,ANULZD_ACTL_BAL
,ACCT_STATS_CNT
,LVL_NUM
,METHOD_TXT
)
SELECT YYMM_ID
,ALOCN_SRC_CD AS BUS_UNIT_ID
,PROD_LINE_CD AS BUS_UNIT_PROD_LINE_CD
,ALOCN_SRC_CD
,PROD_LINE_CD AS ALOCN_SRC_PROD_LINE_CD
,ALOCN_SRC_CD AS ORIG_ALOCN_SRC_CD
,LINE_ITEM_NUM
,'D' AS BAL_ORIGTN_IND
,FIN_ALOCN_AMT AS ANULZD_ACTL_BAL
,0.0 AS ACCT_STATS_CNT
,0 AS LVL_NUM
,'D-INIT' AS METHOD_TXT
-- FROM MISProcess.FIN_FTP
FROM PUASFIN.FocusResults.BUPALLGE
END
ELSE
BEGIN
DELETE FROM MISWork.AB
WHERE LVL_NUM >= @MinLevel
END
DECLARE @LVL_NUM AS int
SET @LVL_NUM = @MinLevel
WHILE @LVL_NUM <= @MaxLevel
BEGIN
DECLARE @LevelStart AS varchar(50)
SET @LevelStart = 'Level:' + CONVERT(varchar, @LVL_NUM)
RAISERROR (@LevelStart, 10, 1) WITH NOWAIT
RAISERROR ('STD_COST_PCT allocations - No D - B records', 10, 1) WITH NOWAIT
-- STD_COST_PCT allocations - No D - B records
INSERT INTO MISWork.AB (
YYMM_ID
,BUS_UNIT_ID
,BUS_UNIT_PROD_LINE_CD
,ALOCN_SRC_CD
,ALOCN_SRC_PROD_LINE_CD
,ORIG_ALOCN_SRC_CD
,LINE_ITEM_NUM
,BAL_ORIGTN_IND
,ANULZD_ACTL_BAL
,ACCT_STATS_CNT
,LVL_NUM
,METHOD_TXT
)
SELECT ALOCNS.YYMM_ID
,ALOCNS.BUS_UNIT_ID
,ALOCNS.BUS_UNIT_PROD_LINE_CD
,ALOCNS.BUS_UNIT_ID AS ALOCN_SRC_CD
,ALOCNS.BUS_UNIT_PROD_LINE_CD AS ALOCN_SRC_PROD_LINE_CD
,ALOCNS.BUS_UNIT_ID AS ORIG_ALOCN_SRC_CD
,ALOCNS.LINE_ITEM_NUM
,'B' AS BAL_ORIGTN_IND
,-1.0 * ROUND(ALOCNS.ANULZD_ACTL_BAL, 2) AS ANULZD_ACTL_BAL
,ROUND(ALOCNS.ACCT_STATS_CNT, 2) AS ACCT_STATS_CNT
,@LVL_NUM AS LVL_NUM
,'NO-D-B' AS METHOD_TXT
FROM [MISProcess].[udf_MR_ALOCNS_STD_COST_PCT_NO_D](@DATA_DT_ID_use, @LVL_NUM)
AS ALOCNS
RAISERROR ('STD_COST_PCT allocations - No D - A records', 10, 1) WITH NOWAIT
-- STD_COST_PCT allocations - No D - A records
INSERT INTO MISWork.AB (
YYMM_ID
,BUS_UNIT_ID
,BUS_UNIT_PROD_LINE_CD
,ALOCN_SRC_CD
,ALOCN_SRC_PROD_LINE_CD
,ORIG_ALOCN_SRC_CD
,LINE_ITEM_NUM
,BAL_ORIGTN_IND
,ANULZD_ACTL_BAL
,ACCT_STATS_CNT
,LVL_NUM
,METHOD_TXT
)
SELECT ALOCNS.YYMM_ID
,BLOCK.TO_ALOCN_SRC_CD AS BUS_UNIT_ID
,ALOCNS.ALOCN_SRC_PROD_LINE_CD AS BUS_UNIT_PROD_LINE_CD
,ALOCNS.ALOCN_SRC_CD
,ALOCNS.BUS_UNIT_PROD_LINE_CD AS ALOCN_SRC_PROD_LINE_CD
,ALOCNS.ORIG_ALOCN_SRC_CD
,ALOCNS.LINE_ITEM_NUM
,'A' AS BAL_ORIGTN_IND
,ROUND(ALOCNS.ANULZD_ACTL_BAL, 2) AS ANULZD_ACTL_BAL
,ROUND(ALOCNS.ACCT_STATS_CNT, 2) AS ACCT_STATS_CNT
,@LVL_NUM AS LVL_NUM
,'NO-D-A' AS METHOD_TXT
FROM [MISProcess].[udf_MR_ALOCNS_STD_COST_PCT_NO_D](@DATA_DT_ID_use, @LVL_NUM)
AS ALOCNS
INNER JOIN MISProcess.LKP_BLOCK AS BLOCK
-- TODO: Can this be moved into the udf above?
ON BLOCK.DATA_DT_ID = @DATA_DT_ID_use
AND BLOCK.FROM_ALOCN_SRC_CD = ALOCNS.BUS_UNIT_ID
RAISERROR ('STD_COST_PCT allocations - B records', 10, 1) WITH NOWAIT
-- STD_COST_PCT allocations - B records
INSERT INTO MISWork.AB (
YYMM_ID
,BUS_UNIT_ID
,BUS_UNIT_PROD_LINE_CD
,ALOCN_SRC_CD
,ALOCN_SRC_PROD_LINE_CD
,ORIG_ALOCN_SRC_CD
,LINE_ITEM_NUM
,BAL_ORIGTN_IND
,ANULZD_ACTL_BAL
,ACCT_STATS_CNT
,LVL_NUM
,METHOD_TXT
)
SELECT ALOCNS.YYMM_ID
,ALOCNS.BUS_UNIT_ID
,ALOCNS.BUS_UNIT_PROD_LINE_CD
,ALOCNS.ALOCN_SRC_CD
,ALOCNS.BUS_UNIT_PROD_LINE_CD AS ALOCN_SRC_PROD_LINE_CD
,ALOCNS.ORIG_ALOCN_SRC_CD
,ALOCNS.LINE_ITEM_NUM
,'B' AS BAL_ORIGTN_IND
,-1.0 * ROUND(ALOCNS.ANULZD_ACTL_BAL * RATios.RATIO, 2) AS ANULZD_ACTL_BAL
,ROUND(ALOCNS.ACCT_STATS_CNT, 2) AS ACCT_STATS_CNT
,@LVL_NUM AS LVL_NUM
,'STD-B' AS METHOD_TXT
FROM [MISProcess].[udf_MR_ALOCNS_STD_COST_PCT](@DATA_DT_ID_use, @LVL_NUM)
AS ALOCNS
INNER JOIN [MISProcess].[udf_MR_RATIOS_STD_COST_PCT](@DATA_DT_ID_use, @LVL_NUM)
AS RATIOS
ON RATIOS.YYMM_ID = ALOCNS.YYMM_ID
AND RATIOS.BUS_UNIT_ID = ALOCNS.BUS_UNIT_ID
AND RATIOS.LINE_ITEM_NUM = ALOCNS.LINE_ITEM_NUM
RAISERROR ('STD_COST_PCT allocations - A records', 10, 1) WITH NOWAIT
-- STD_COST_PCT allocations - A records
;
WITH CORRECTED_ALOCNS
AS (
SELECT ALOCNS.YYMM_ID
,ALOCNS.BUS_UNIT_ID
,ALOCNS.BUS_UNIT_PROD_LINE_CD
,ALOCNS.ALOCN_SRC_CD
,ALOCNS.ALOCN_SRC_PROD_LINE_CD
,ALOCNS.ORIG_ALOCN_SRC_CD
,ALOCNS.LINE_ITEM_NUM
,ALOCNS.ANULZD_ACTL_BAL * RATIOS.RATIO AS ANULZD_ACTL_BAL
,CASE WHEN RATIOS.RATIO <> 1.0
THEN RATIOS.RATIO
ELSE ALOCNS.ACCT_STATS_CNT
END AS ACCT_STATS_CNT
FROM [MISProcess].[udf_MR_CORR_ALOCNS_STD_COST_PCT](@DATA_DT_ID_use, @LVL_NUM)
AS ALOCNS
INNER JOIN [MISProcess].[udf_MR_RATIOS_STD_COST_PCT](@DATA_DT_ID_use, @LVL_NUM)
AS RATIOS
ON RATIOS.YYMM_ID = ALOCNS.YYMM_ID
AND RATIOS.BUS_UNIT_ID = ALOCNS.ALOCN_SRC_CD
AND RATIOS.LINE_ITEM_NUM = ALOCNS.LINE_ITEM_NUM
),
ROUNDED_ALOCNS
AS (
SELECT YYMM_ID
,BUS_UNIT_ID
,BUS_UNIT_PROD_LINE_CD
,ALOCN_SRC_CD
,ALOCN_SRC_PROD_LINE_CD
,ORIG_ALOCN_SRC_CD
,LINE_ITEM_NUM
,CASE WHEN ABS(ANULZD_ACTL_BAL) < 0.05 THEN 0.0
WHEN ABS(ANULZD_ACTL_BAL) > 0.05
AND ABS(ANULZD_ACTL_BAL) < 0.10
THEN 0.10 * SIGN(ANULZD_ACTL_BAL)
ELSE ANULZD_ACTL_BAL
END AS ANULZD_ACTL_BAL
,ACCT_STATS_CNT
FROM CORRECTED_ALOCNS
)
INSERT INTO MISWork.AB (
YYMM_ID
,BUS_UNIT_ID
,BUS_UNIT_PROD_LINE_CD
,ALOCN_SRC_CD
,ALOCN_SRC_PROD_LINE_CD
,ORIG_ALOCN_SRC_CD
,LINE_ITEM_NUM
,BAL_ORIGTN_IND
,ANULZD_ACTL_BAL
,ACCT_STATS_CNT
,LVL_NUM
,METHOD_TXT
)
SELECT YYMM_ID
,BUS_UNIT_ID
,BUS_UNIT_PROD_LINE_CD
,ALOCN_SRC_CD
,ALOCN_SRC_PROD_LINE_CD
,ORIG_ALOCN_SRC_CD
,LINE_ITEM_NUM
,'A' AS BAL_ORIGTN_IND
,ROUND(ANULZD_ACTL_BAL, 2) AS ANULZD_ACTL_BAL
,ROUND(ACCT_STATS_CNT, 2) AS ACCT_STATS_CNT
,@LVL_NUM AS LVL_NUM
,'STD-A' AS METHOD_TXT
FROM ROUNDED_ALOCNS
WHERE ANULZD_ACTL_BAL <> 0.0
OR ACCT_STATS_CNT <> 0.0
RAISERROR ('COLLAPSE, BLOCK 100 ALOCN_PCT - B records', 10, 1) WITH NOWAIT
-- COLLAPSE, BLOCK 100% ALOCN_PCT - B records
INSERT INTO MISWork.AB (
YYMM_ID
,BUS_UNIT_ID
,BUS_UNIT_PROD_LINE_CD
,ALOCN_SRC_CD
,ALOCN_SRC_PROD_LINE_CD
,ORIG_ALOCN_SRC_CD
,LINE_ITEM_NUM
,BAL_ORIGTN_IND
,ANULZD_ACTL_BAL
,ACCT_STATS_CNT
,LVL_NUM
,METHOD_TXT
)
SELECT BALANCES.YYMM_ID
,BALANCES.BUS_UNIT_ID
,BALANCES.BUS_UNIT_PROD_LINE_CD
,BALANCES.BUS_UNIT_ID AS ALOCN_SRC_CD
,BALANCES.BUS_UNIT_PROD_LINE_CD AS ALOCN_SRC_PROD_LINE_CD
,BALANCES.ORIG_ALOCN_SRC_CD
,BALANCES.ALOCN_LINE_ITEM_NUM AS LINE_ITEM_NUM
,'B' AS BAL_ORIGTN_IND
,-1.0 * BALANCES.ANULZD_ACTL_BAL
,ALOCN_SUMRY.ACCT_STATS_CNT
,@LVL_NUM AS LVL_NUM
,'BLOCK-100' AS METHOD_TXT
FROM [MISProcess].[udf_MR_BALANCES_BLOCK_100_PCT](@DATA_DT_ID_use, @LVL_NUM)
AS BALANCES
INNER JOIN [MISProcess].[udf_MR_ALOCN_SUMRY_BLOCK_100_PCT](@DATA_DT_ID_use, @LVL_NUM)
AS ALOCN_SUMRY
ON ALOCN_SUMRY.YYMM_ID = BALANCES.YYMM_ID
AND ALOCN_SUMRY.BUS_UNIT_ID = BALANCES.BUS_UNIT_ID
AND ALOCN_SUMRY.BUS_UNIT_PROD_LINE_CD = BALANCES.BUS_UNIT_PROD_LINE_CD
AND ALOCN_SUMRY.ALOCN_SRC_CD = BALANCES.ALOCN_SRC_CD
AND ALOCN_SUMRY.ALOCN_SRC_PROD_LINE_CD = BALANCES.ALOCN_SRC_PROD_LINE_CD
AND ALOCN_SUMRY.ORIG_ALOCN_SRC_CD = BALANCES.ORIG_ALOCN_SRC_CD
RAISERROR ('COLLAPSE, BLOCK 100 ALOCN_PCT - A records', 10, 1) WITH NOWAIT
-- COLLAPSE, BLOCK 100% ALOCN_PCT - A records
INSERT INTO MISWork.AB (
YYMM_ID
,BUS_UNIT_ID
,BUS_UNIT_PROD_LINE_CD
,ALOCN_SRC_CD
,ALOCN_SRC_PROD_LINE_CD
【问题讨论】：
看起来我在 SO 上遇到了问题大小的限制，因此对我给出的特定匹配的 UDF 集的实际调用丢失了，但这些调用与确实显示的调用相似起来。 【参考方案1】：是的，重新编译应该扩展到 TV UDFS。
但是，我会使用参数掩码而不是 RECOMPILE。
-
使用这样的查询，编译成本会很高
当 UDF 未嵌套时，参数掩码也将应用。 TV UDF 没有这样的计划：它们是调用查询的一部分，因为它们是非嵌套的。
您能否将一些 UDF 调用分解为临时表，然后加入临时表？ 我敢打赌，当 UDF 未嵌套时，查询太复杂而无法有效运行。优化者可能需要一周的时间才能找到包含如此复杂事物的理想计划。使用临时表（不是表变量），我想你会得到可观的改进。
我自己在一些较大的查询中使用了这种技术（为金融工具生成定价树）
150,000 行的事实被我认为的绝对复杂性所掩盖。
编辑：
TVF 不需要参数屏蔽，因为它们只是宏。您可以将其替换为 CTE 或派生表。
在这里查看我的答案：Does query plan optimizer works well with joined/filtered table-valued functions 还有Tony Rogerson on Views
【讨论】：
我假设您的意思是 SP 中的参数屏蔽。因为当我最初查看它时，我找不到在内联 TVF 中进行参数屏蔽的方法。是的，我对每个 SP 都这样做。我希望我们使用 2008，这样我们就可以使用 OPTIMIZE FOR UNKNOWN 选项。 您不需要屏蔽 TVF，因为它不是离散对象。它已扩展为包含查询。更新答案【参考方案2】：你能发布实际的 T-SQL 而不是伪 SQL 吗？ 您所描述的听起来像是随着@LVL_NUM 的增加而对越来越大的结果集进行表扫描，并且可能 RECOMPILE 在那里无济于事。但是，基于伪SQL，除了伪猜测之外真的很难给出任何东西......
【讨论】：
似乎测试用例在帖子中被截断，并且从代码本身来看，udf_MR_ALSRC 丢失了。您能否提供参与者桌子的桌子大小的大致情况？ 另外，我看到您尝试添加 SCHEMABINDING，为什么要删除它？ 在增加@LVL_NUM 时，您是否比较了查询的实际执行计划？如果您的理论是正确的，那么计划将改变用坏访问路径（扫描）替换好的访问路径（搜索）。如果我的理论是正确的，那么计划将保持不变，但执行的权重会因@LVL_NUM 较高的重复执行而呈指数级增加（您会看到计划中“执行次数”不断增加的元素）。 如果我一次运行一个级别，执行计划似乎没有改变。这让我相信 UDF 有某种缓存计划。当然，在循环中运行时，每次迭代都使用相同的计划，所以这可能是根本原因。如果我删除并重新创建 UDF，我会注意到性能差异，因此那里似乎有一些缓存。 如果在冷缓存 (DBCC FREESYTEMCACHE) 上直接以高 @LVL_NUM 开始，您是否有更好的计划？此外，从缓存中删除计划后，执行时间和编译时间是多少？ SET STATISTICS TIME ON 获取编译时间。【参考方案3】：使用表值函数（尤其是多语句表值函数）要注意的一件事是生成的表，就像表变量一样，没有列统计信息和索引。
我倾向于谨慎使用 TVF。
【讨论】：
即使在第 6 级，该函数返回的表也很小——我认为是 150,000 行。 而且这些都是内联表值函数。 我理解这基本上被视为参数化视图 - 即优化器可以折叠它们并将它们推到一起，就像它们被内联写出一样。以上是关于T-SQL 流程设计和执行计划（UDF 参数嗅探？）的主要内容，如果未能解决你的问题，请参考以下文章
理解性能的奥秘——应用程序中慢，SSMS中快——案例：如何应对参数嗅探
理解性能的奥秘——应用程序中慢，SSMS中快——案例：如何应对参数嗅探