Cannot use UPDATE with OUTPUT clause when a trigger is on the table

问题

I'm performing an UPDATE with OUTPUT query:

UPDATE BatchReports
SET IsProcessed = 1
OUTPUT inserted.BatchFileXml, inserted.ResponseFileXml, deleted.ProcessedDate
WHERE BatchReports.BatchReportGUID = @someGuid

This statement is well and fine; until a trigger is defined on the table. Then my UPDATE statement will get the error 334:

The target table 'BatchReports' of the DML statement cannot have any enabled triggers if the statement contains an OUTPUT clause without INTO clause

Now this problem is explained in a blog post by the SQL Server team:

The error message is self-explanatory

And they also give solutions:

The application was changed to utilize the INTO clause

Except I cannot make heads or tails of the entirety of the blog post.

So let me ask my question: What should i change my UPDATE to so that it works?

See also

• UPDATE with OUTPUT clause – Triggers – and SQLMoreResults
• Why is the target table of a MERGE statement not allowed to have enabled rules?
• Statement contains an OUTPUT clause without INTO clause error

回答1:

To work around this restriction you need to OUTPUT INTO ... something. e.g. declare an intermediary table variable to be the target then SELECT from that.

DECLARE @T TABLE (
  BatchFileXml    XML,
  ResponseFileXml XML,
  ProcessedDate   DATE,
  RowVersion      BINARY(8) )

UPDATE BatchReports
SET    IsProcessed = 1
OUTPUT inserted.BatchFileXml,
       inserted.ResponseFileXml,
       deleted.ProcessedDate,
       inserted.Timestamp
INTO @T
WHERE  BatchReports.BatchReportGUID = @someGuid

SELECT *
FROM   @T 

As cautioned in the other answer if your trigger writes back to the rows modified by the UPDATE statement itself in such a way that it affects the columns that you are OUTPUT-ing then you may not find the results useful but this is only a subset of triggers. The above technique works fine in other cases, such as triggers recording to other tables for audit purposes, or returning inserted identity values even if the original row is written back to in the trigger.

回答2:

Visibility Warning: Don't use the highest voted answer. It will give incorrect values. Read on for the way it's wrong.

---

Given the kludge needed to make UPDATE with OUTPUT work in SQL Server 2008 R2, i changed my query from:

UPDATE BatchReports  
SET IsProcessed = 1
OUTPUT inserted.BatchFileXml, inserted.ResponseFileXml, deleted.ProcessedDate
WHERE BatchReports.BatchReportGUID = @someGuid

to:

SELECT BatchFileXml, ResponseFileXml, ProcessedDate FROM BatchReports
WHERE BatchReports.BatchReportGUID = @someGuid

UPDATE BatchReports
SET IsProcessed = 1
WHERE BatchReports.BatchReportGUID = @someGuid

Basically i stopped using OUTPUT. This isn't so bad as Entity Framework itself uses this very same hack!

Hopefully 2012 2014 2016 2018 will have a better implementation.

---

Update: using OUTPUT is harmful

The problem we started with was trying to use the OUTPUT clause to retrieve the "after" values in a table:

UPDATE BatchReports
SET IsProcessed = 1
OUTPUT inserted.LastModifiedDate, inserted.RowVersion, inserted.BatchReportID
WHERE BatchReports.BatchReportGUID = @someGuid

That then hits the well-know limitation (won't-fix bug) in SQL Server:

The target table 'BatchReports' of the DML statement cannot have any enabled triggers if the statement contains an OUTPUT clause without INTO clause

Workaround Attempt #1

So we try something where we will use an intermediate TABLE variable to hold the OUTPUT results:

DECLARE @t TABLE (
   LastModifiedDate datetime,
   RowVersion timestamp, 
   BatchReportID int
)

UPDATE BatchReports
SET IsProcessed = 1
OUTPUT inserted.LastModifiedDate, inserted.RowVersion, inserted.BatchReportID
INTO @t
WHERE BatchReports.BatchReportGUID = @someGuid

SELECT * FROM @t

Except that fails because you're not allowed to insert a timestamp into the table (even a temporary table variable).

Workaround Attempt #2

We secretly know that a timestamp is actually a 64-bit (aka 8 byte) unsigned integer. We can change our temporary table definition to use binary(8) rather than timestamp:

DECLARE @t TABLE (
   LastModifiedDate datetime,
   RowVersion binary(8), 
   BatchReportID int
)

UPDATE BatchReports
SET IsProcessed = 1
OUTPUT inserted.LastModifiedDate, inserted.RowVersion, inserted.BatchReportID
INTO @t
WHERE BatchReports.BatchReportGUID = @someGuid

SELECT * FROM @t

And that works, except that the value are wrong.

The timestamp RowVersion we return is not the value of the timestamp as it existed after the UPDATE completed:

• returned timestamp: 0x0000000001B71692
• actual timestamp: 0x0000000001B71693

That is because the values OUTPUT into our table are not the values as they were at the end of the UPDATE statement:

• UPDATE statement starting
  • modify row
  • timestamp is updated
  • retrieve new timestamp
  • trigger runs
    • modify row
    • timestamp is updated
• UPDATE statement complete

This means:

• We do not get the timestamp as it exists at the end of the UPDATE statement
• we get the timestamp as it was in the indeterminate middle of the UPDATE statement
• we do not get the correct timestamp

The same is true of any trigger that modifies any value in the row. The OUTPUT will not OUTPUT the value as of the end of the UPDATE.

This means you not trust OUTPUT to return any correct values

This painful reality is documented in the BOL:

Columns returned from OUTPUT reflect the data as it is after the INSERT, UPDATE, or DELETE statement has completed but before triggers are executed.

How did Entity Framework solve it?

The .NET Entity Framework uses rowversion for Optimistic Concurrency. The EF depends on knowing the value of the timestamp as it after they issue an UPDATE.

Since you cannot use OUTPUT for any important data, Microsoft's Entity Framework uses the same workaround that i do:

Workaround #3 - Final

In order to retrieve the after values, Entity Framework issues:

UPDATE [dbo].[BatchReports]
SET [IsProcessed] = @0
WHERE (([BatchReportGUID] = @1) AND ([RowVersion] = @2))

SELECT [RowVersion], [LastModifiedDate]
FROM [dbo].[BatchReports]
WHERE @@ROWCOUNT > 0 AND [BatchReportGUID] = @1

Don't use OUTPUT.

Yes it suffers from a race condition, but that's the best SQL Server can do.

What about INSERTs

Do what Entity Framework does:

SET NOCOUNT ON;

DECLARE @generated_keys table([CustomerID] int)

INSERT Customers (FirstName, LastName)
OUTPUT inserted.[CustomerID] INTO @generated_keys
VALUES ('Steve', 'Brown')

SELECT t.[CustomerID], t.[CustomerGuid], t.[RowVersion], t.[CreatedDate]
FROM @generated_keys AS g
   INNER JOIN Customers AS t
   ON g.[CustomerGUID] = t.[CustomerGUID]
WHERE @@ROWCOUNT > 0

回答3:

Why put all needed columns into table variable? We just need primary key and we can read all data after the UPDATE. There is no race when you use transaction:

DECLARE @t TABLE (ID INT PRIMARY KEY);

BEGIN TRAN;

UPDATE BatchReports SET 
    IsProcessed = 1
OUTPUT inserted.ID INTO @t(ID)
WHERE BatchReports.BatchReportGUID = @someGuid;

SELECT b.* 
FROM @t t JOIN BatchReports b ON t.ID = b.ID;

COMMIT;

来源：https://stackoverflow.com/questions/13198476/cannot-use-update-with-output-clause-when-a-trigger-is-on-the-table