Nested stored procedures containing TRY CATCH ROLLBACK pattern? - sql-server-2005

Question

I'm interested in the side effects and potential problems of the following pattern:
CREATE PROCEDURE [Name]
AS
BEGIN
BEGIN TRANSACTION
BEGIN TRY
[...Perform work, call nested procedures...]
END TRY
BEGIN CATCH
ROLLBACK TRANSACTION
RAISERROR [rethrow caught error using #ErrorNumber, #ErrorMessage, etc]
END CATCH
END
To the best of my understanding this pattern is sound when used with a single procedure - the procedure will either complete all of its statements without error, or it will rollback all actions and report the error.
However when one stored procedure calls another stored procedure to do some sub-unit of work (with the understanding that the smaller procedure is sometimes called on its own) I see an issue coming about with relation to rollbacks - an informational message (Level 16) is issued stating The ROLLBACK TRANSACTION request has no corresponding BEGIN TRANSACTION.. This I assume is because the rollback in the sub-procedure is always rolling back the outer-most transaction, not just the transaction started in the sub-procedure.
I do want the whole thing rolled back and aborted if any error occurs (and the error reported to the client as an SQL error), I'm just not sure of all the side effects that come from the outer layers trying to rollback a transaction that has already been rolled back. Perhaps a check of ##TRANCOUNT before doing a rollback at each TRY CATCH layer?
Finally there is the client end (Linq2SQL), which has it's own transaction layer:
try
{
var context = new MyDataContext();
using (var transaction = new TransactionScope())
{
// Some Linq stuff
context.SubmitChanges();
context.MyStoredProcedure();
transactionComplete();
}
}
catch
{
// An error occured!
}
In the event that a stored procedure, "MySubProcedure", called inside MyStoredProcedure raises an error, can I be sure that everything previously done in MyStoredProcedure will be rolled back, all the Linq operations made by SubmitChanges will be rolled back, and finally that the error will be logged? Or what do I need to change in my pattern to ensure the whole operation is atomic, while still allowing the child parts to be used individually (i.e. the sub-procedures should still have the same atomic protection)

Answer 1

This is our template (error logging removed)
This is designed to handle
Paul Randal's article "No such thing as a nested transaction in SQL Server"
Error 266
Trigger Rollbacks
Explanations:
all TXN begin and commit/rollbacks must be paired so that ##TRANCOUNT is the same on entry and exit
mismatches of ##TRANCOUNT cause error 266 because
BEGIN TRAN increments ##TRANCOUNT
COMMIT decrements ##TRANCOUNT
ROLLBACK returns ##TRANCOUNT to zero
You can not decrement ##TRANCOUNT for the current scope
This is what you'd think is the "inner transaction"
SET XACT_ABORT ON suppresses error 266 caused by mismatched ##TRANCOUNT
And also deals with issues like this "SQL Server Transaction Timeout" on dba.se
This allows for client side TXNs (like LINQ)
A single stored procedure may be part of distributed or XA transaction, or simply one initiated in client code (say .net TransactionScope)
Usage:
Each stored proc must conform to the same template
Summary
So don't create more TXNs than you need
The code
CREATE PROCEDURE [Name]
AS
SET XACT_ABORT, NOCOUNT ON
DECLARE #starttrancount int
BEGIN TRY
SELECT #starttrancount = ##TRANCOUNT
IF #starttrancount = 0
BEGIN TRANSACTION
[...Perform work, call nested procedures...]
IF #starttrancount = 0
COMMIT TRANSACTION
END TRY
BEGIN CATCH
IF XACT_STATE() <> 0 AND #starttrancount = 0
ROLLBACK TRANSACTION;
THROW;
--before SQL Server 2012 use
--RAISERROR [rethrow caught error using #ErrorNumber, #ErrorMessage, etc]
END CATCH
GO
Notes:
The rollback check is actually redundant because of SET XACT_ABORT ON. However, it makes me feel better, looks odd without, and allows for situations where you don't want it on
Remus Rusanu has a similar shell that uses save points. I prefer an atomic DB call and don't use partial updates like their article

Answer 2

I am not a Linq guy (and neither is Erland), but he wrote the absolute bibles on error handling. Outside of the complications Linq might add to your problem, all of your other questions should be answered here:
http://www.sommarskog.se/error_handling/Part1.html
(Old link: http://www.sommarskog.se/error_handling_2005.html)

Answer 3

To solve the issue of returning the error number and line number mentioned by #AlexKuznetsov, one can raise the error as such:
DECLARE #ErrorMessage NVARCHAR(4000)
DECLARE #ErrorSeverity INT
DECLARE #ErrorState INT
DECLARE #ErrorLine INT
DECLARE #ErrorNumber INT
SELECT #ErrorMessage = ERROR_MESSAGE(),
#ErrorSeverity = ERROR_SEVERITY(),
#ErrorState = ERROR_STATE(),
#ErrorNumber = ERROR_NUMBER(),
#ErrorLine = ERROR_LINE()
RAISERROR (#ErrorMessage, #ErrorSeverity, #ErrorState, #ErrorNumber, #ErrorLine)

Answer 4

-- #Amanda method above doesnt return correct error number
DECLARE
#ErrorMessage nvarchar(4000),
#ErrorSeverity int,
#ErrorState int,
#ErrorLine int,
#ErrorNumber int
BEGIN TRY
SELECT 1/0; -- CATCH me
END TRY
BEGIN CATCH
DECLARE #err int = ##ERROR
PRINT #err -- 8134, divide by zero
PRINT ERROR_NUMBER() -- 8134
SELECT
#ErrorMessage = ERROR_MESSAGE(),
#ErrorSeverity = ERROR_SEVERITY(),
#ErrorState = ERROR_STATE(),
#ErrorNumber = ERROR_NUMBER(),
#ErrorLine = ERROR_LINE()
-- error number = 50000 :(
RAISERROR (#ErrorMessage, #ErrorSeverity, #ErrorState, #ErrorNumber, #ErrorLine)
END CATCH
-- error number = 8134
SELECT 1/0

Answer 5

In case no special error handling needed in CATCH except rethrow and stored procs call chain isn't too long it may be suitable to use such simple template:
create procedure someNestedSP
as
SET XACT_ABORT ON
begin transaction
-- do some work or call some other similar SP
commit transaction
It would also rollback root transaction with all "nested" ones in case of any error but the code is shorter and more straightforward than #gbn's solution. Still XACT_ABORT takes care of most issues mentioned there.
There may be addiotional overhead for transaction nesting but it may be not too high, I guess.