我有MERGE
查询在我的数据库上创建一个upsert操作,并通过我的应用程序输入数据。当我打电话给它以保存我在大型交易(> 5000)上的数据时,它需要非常很长时间(~20-40秒)。
这是我的MERGE
声明
MERGE TestTable AS target USING (SELECT @Guid) AS source (target.Guid = source.Guid)
WHEN MATCHED THEN
UPDATE TestTable SET Column1 = @Column1, Column2 = @Column2 WHERE Guid = @Guid
WHEN NOT MATCHED THEN
INSERT INTO TestTable (Column1, Column2) VALUES (@Column1, @Column2)
OUTPUT $action
我在.NET代码中一次在一个对象上调用它。
我在SQL Express 2008活动监视器的活动监视器中注意到,由于调用查询的所有不同参数排列,计划计数最多可达900左右。我还注意到,如果我在不久之后使用稍微不同的参数重复相同的保存,它会更快地保存很多(~2秒)。
这是潜在的性能问题以及长时间保存的原因吗?
我正在使用SQL Express 2008 R2。
编辑:这是计划:
|--Compute Scalar(DEFINE:([Expr1044]=CASE WHEN [Action1004]=(1) THEN N'UPDATE' ELSE CASE WHEN [Action1004]=(4) THEN N'INSERT' ELSE N'DELETE' END END))
|--Assert(WHERE:(CASE WHEN NOT [Pass1238] AND [Expr1237] IS NULL THEN (0) ELSE NULL END))
|--Nested Loops(Left Semi Join, PASSTHRU:([Action1004]=(3) OR [C:\DATABASE.MDF].[dbo].[DoorTable].[CarTable_Guid] as [target].[CarTable_Guid] IS NULL), OUTER REFERENCES:([C:\DATABASE.MDF].[dbo].[DoorTable].[CarTable_Guid]), DEFINE:([Expr1237] = [PROBE VALUE]))
|--Clustered Index Merge(OBJECT:([C:\DATABASE.MDF].[dbo].[DoorTable].[DoorTable_PK] AS [target]), OBJECT:([C:\DATABASE.MDF].[dbo].[DoorTable].[DoorTable_FI01] AS [target]), SET:(Insert, [C:\DATABASE.MDF].[dbo].[DoorTable].[Column1] as [target].[Column1] = [Expr1005],[C:\DATABASE.MDF].[dbo].[DoorTable].[Column2] as [target].[Column2] = [Expr1006],[C:\DATABASE.MDF].[dbo].[DoorTable].[Column3] as [target].[Column3] = [Expr1007],[C:\DATABASE.MDF].[dbo].[DoorTable].[Column4] as [target].[Column4] = [Expr1008],[C:\DATABASE.MDF].[dbo].[DoorTable].[Column5] as [target].[Column5] = [Expr1009],[C:\DATABASE.MDF].[dbo].[DoorTable].[Column6] as [target].[Column6] = [Expr1010],[C:\DATABASE.MDF].[dbo].[DoorTable].[Column7] as [target].[Column7] = [Expr1011],[C:\DATABASE.MDF].[dbo].[DoorTable].[Column8] as [target].[Column8] = [Expr1012],...
| |--Compute Scalar(DEFINE:([Action1004]=[Action1004], [Expr1198]=[Expr1198]))
| |--Top(TOP EXPRESSION:((1)))
| |--Compute Scalar(DEFINE:([Expr1198]=CASE WHEN [Action1004] = (1) THEN CASE WHEN [Expr1099] THEN (0) ELSE (1) END ELSE [Action1004] END))
| |--Compute Scalar(DEFINE:([Expr1005]=CONVERT_IMPLICIT(nvarchar(64),[@Column1],0), [Expr1006]=CONVERT_IMPLICIT(nvarchar(64),[@Column2],0), [Expr1007]=CONVERT_IMPLICIT(nvarchar(64),[@Column3],0), [Expr1008]=[@Column4], [Expr1009]=CONVERT_IMPLICIT(nvarchar(64),[@Column5],0), [Expr1010]=[@Column6], [Expr1011]=[@Column7], [Expr1012]=CONVERT_IMPLICIT(float(53),[@Column8],0),[Expr1099]=[Action1004]=(1) AND CASE WHEN [C:\DATABASE.MDF].[dbo].[DoorTable].[CarTable_Guid] as [target].[CarTable_Guid] = CONVERT_IMPLICIT(nvarchar(32),[@CarTable_Guid],0) THEN (1) ELSE (0) END))
| |--Compute Scalar(DEFINE:([Action1004]=ForceOrder(CASE WHEN [TrgPrb1002] IS NOT NULL THEN (1) ELSE (4) END)))
| |--Nested Loops(Left Outer Join)
| |--Constant Scan
| |--Compute Scalar(DEFINE:([TrgPrb1002]=(1)))
| |--Clustered Index Seek(OBJECT:([C:\DATABASE.MDF].[dbo].[DoorTable].[DoorTable_PK] AS [target]), SEEK:([target].[Guid]=CONVERT_IMPLICIT(nvarchar(1),[@Guid],0)) ORDERED FORWARD)
|--Clustered Index Seek(OBJECT:([C:\DATABASE.MDF].[dbo].[CarTable].[CarTable_PK]), SEEK:([C:\DATABASE.MDF].[dbo].[CarTable].[Guid]=[C:\DATABASE.MDF].[dbo].[DoorTable].[CarTable_Guid] as [target].[CarTable_Guid]) ORDERED FORWARD)
答案 0 :(得分:1)
不是在循环5000
次中执行此操作,而是将更好的包装放入存储过程中,该过程将TABLE
值作为输入,并执行批量更新:
CREATE TYPE paramTable AS TABLE
(
guid UNIQUEIDENTIFIER NOT NULL PRIMARY KEY,
column1 INT,
column2 VARCHAR(100)
)
CREATE PROCEDURE prcMergeInput(@mytable paramTable)
AS
MERGE TestTable AS target
USING (
SELECT column1, column2, guid
FROM @mytable
) AS source
ON (target.Guid = source.Guid)
WHEN MATCHED THEN
UPDATE TestTable
SET Column1 = source.Column1,
Column2 = source.Column2
WHEN NOT MATCHED THEN
INSERT
INTO TestTable (Column1, Column2)
VALUES (source.Column1, source.Column2)
OUTPUT INSERTED.guid
另外,请确保您在TestTable (guid)
上有索引,或者将其声明为PRIMARY KEY
。
答案 1 :(得分:1)
要验证缓存计划的来源,您可以查询包含所有缓存计划的动态管理视图:
SELECT text
FROM sys.dm_exec_cached_plans
CROSS APPLY
sys.dm_exec_sql_text(plan_handle)
WHERE text LIKE ‘%SnippetFromYourQuery%’
虽然看起来您正在对查询进行正确的参数化,但您可以通过启用forced parameterization来测试它:
alter database YourDb forced
如果这会减少运行时间,则应调查查询的哪一部分包含硬编码值而不是参数。 SQL事件探查器应该很容易。
答案 2 :(得分:1)
我已经解决了我遇到的问题。让我弄明白的原因是看看这些家伙遇到的问题:nHibernate Recompiles and Execution Plans
基本上,在.NET代码中,它没有定义DbParameter.Size属性。由于不同的参数大小会导致创建不同的执行计划,因此我所有参数的每个排列都会导致创建和缓存不同的计划。
我所要做的就是从我的DDL脚本中将DbParameter.Size设置为每个相应列的大小!哇。