Commiting only Specific Changes made inside a TRANSACTION which may ROLLBACK

Question

This is a significant edit from the original question, making it more concise and covering the points raised by existing answers...

Is it possible to have mulitple changes made to multiple tables, inside a single transaction, and rollback only some of the changes?

In the TSQL below, I would NOT want any of the changes made by "myLogSP" to ever be rolled back. But all changes made by the various myBusinessSPs should rollback if necessary.

BEGIN TRANSACTION  

    EXEC myLogSP

    EXEC @err = myBusinessSPa
    IF (@err <> 0) BEGIN ROLLBACK TRANSACTION RETURN -1 END

    EXEC myLogSP

    EXEC @err = myBusinessSPb
    IF (@err <> 0) BEGIN ROLLBACK TRANSACTION RETURN -1 END

    EXEC myLogSP

    EXEC @err = myBusinessSPc
    IF (@err <> 0) BEGIN ROLLBACK TRANSACTION RETURN -1 END

    EXEC myLogSP

COMMIT TRANSACTION
RETURN 0

The order is important, the myLogSPs must happen between and after the myBusinessSPs (the myLogSPs pick up on the changes made by the myBusinessSPs)

It is also important that all the myBusinessSPs happen inside one transaction to maintain database integrity, and allow all their changes to rollback if necessary.

It's as if I want the myLogSPs to behave as if they're not part of the transaction. It is just an inconvenient fact that they happen to be inside one (by virtue of needing to be called between the myBusinessSPs.)

EDIT:

Final answer is "no", the only option is to redesign the code. Either to using table variables for the logging (as variables don't get rolled back) or redesign the business logic to Not require Transactions...

Answer 1

Use SAVEPOINTs, e.g.

BEGIN TRANSACTION  

    EXEC myLogSP

    SAVE TRANSACTION savepointA
    EXEC @err = myBusinessSPa
    IF (@err <> 0) BEGIN
        ROLLBACK TRANSACTION savepointA
        COMMIT
        RETURN -1
    END

    EXEC myLogSP

    SAVE TRANSACTION savepointB
    EXEC @err = myBusinessSPb
    IF (@err <> 0) BEGIN
        ROLLBACK TRANSACTION savepointB
        COMMIT
        RETURN -1
    END

    EXEC myLogSP

    SAVE TRANSACTION savepointC
    EXEC @err = myBusinessSPc
    IF (@err <> 0) BEGIN
        ROLLBACK TRANSACTION savepointC
        COMMIT
        RETURN -1
    END

    EXEC myLogSP

COMMIT TRANSACTION

---

EDIT

Based on the information provided so far (and my understanding of it) it appears that you will have to re-engineer you logging SPs, either to use variables, or to use files, or to allow them to run 'after the fact' as follows:

BEGIN TRANSACTION  

    SAVE TRANSACTION savepointA
    EXEC @err = myBusinessSPa
    IF (@err <> 0) BEGIN
        ROLLBACK TRANSACTION savepointA
        EXEC myLogSPA -- the call to myBusinessSPa was attempted/failed
        COMMIT
        RETURN -1
    END

    SAVE TRANSACTION savepointB
    EXEC @err = myBusinessSPb
    IF (@err <> 0) BEGIN
        ROLLBACK TRANSACTION savepointB
        EXEC myLogSPA -- the call to myBusinessSPa originally succeeded
        EXEC myLogSPB -- the call to myBusinessSPb was attempted/failed
        COMMIT
        RETURN -1
    END

    SAVE TRANSACTION savepointC
    EXEC @err = myBusinessSPc
    IF (@err <> 0) BEGIN
        ROLLBACK TRANSACTION savepointC
        EXEC myLogSPA -- the call to myBusinessSPa originally succeeded
        EXEC myLogSPB -- the call to myBusinessSPb originally succeeded
        EXEC myLogSPC -- the call to myBusinessSPc was attempted/failed
        COMMIT
        RETURN -1
    END

    EXEC myLogSPA -- the call to myBusinessSPa succeeded
    EXEC myLogSPB -- the call to myBusinessSPb succeeded
    EXEC myLogSPC -- the call to myBusinessSPc succeeded

COMMIT TRANSACTION

Answer 2

You need to basically jump outside of the current context. There are a couple of ways to do that. One (which I have never tried) is to call the CLR to do the insert.

Perhaps a better way though is using the fact that table variables are not affected by transaction. For example:

CREATE TABLE dbo.Test_Transactions
(
     my_string VARCHAR(20) NOT NULL
)
GO

DECLARE
     @tbl TABLE (my_string VARCHAR(20) NOT NULL)

BEGIN TRANSACTION

INSERT INTO dbo.Test_Transactions (my_string) VALUES ('test point one')

INSERT INTO @tbl (my_string) VALUES ('test point two')

INSERT INTO dbo.Test_Transactions (my_string) VALUES ('test point three')

ROLLBACK TRANSACTION

INSERT INTO dbo.Test_Transactions (my_string) select my_string from @tbl

SELECT * FROM dbo.Test_Transactions
SELECT * FROM @tbl
GO

Answer 3

We have had luck with putting the log entries into table variables and then inserting to the real tables after the commit or rollback.

OK if you aren't on SQL Server 2008, then try this method. It's messy and a workaround but it should work. The #temp table and the table variable would have to be set up with the structure of what is returned by the sp.

create table #templog (fie1d1 int, field2 varchar(10))

declare @templog table (fie1d1 int, field2 varchar(10))

BEGIN TRANSACTION      
insert into #templog
Exec my_proc

insert into @templog (fie1d1, field2)
select t.* from #templog t 
left join @templog t2 on t.fie1d1 = t2.fie1d1 where t2.fie1d1 is null

insert into templog
values (1, 'test')

rollback tran
select * from #templog
select * from templog
select * from @templog

Answer 4

Use SAVEPOINTS and TRANSACTION ISOLATION LEVELS.

Answer 5

Wouldn't the easy way be to move the log insertion outside the transaction?

I don't really have an answer for you for the table lock, I think you already have the answer, there will have to be a table lock because the identity column may roll back.

Answer 6

move the BEGIN TRANSACTION statement to after the first insert.

Answer 7

Perhaps you could put the inserts/updates to the business tables in their own atomic transaction t1 and wrap each of these transactions in another transaction t2 that executes the log table update and t1 (the business table updates) without any rollbacks. For example:

BEGIN TRANSACTION t2
     <insert to log>
     <execute stored procedure p1>
END TRANSACTION t2

CREATE PROCEDURE p1
AS
     BEGIN TRANSACTION t1
         <insert to business tables>
         <rollback t1 on error>
     END TRANSACTION t1

I believe that when you rollback t1 in the stored procedure this will leave the calling transaction t2 unaffected.