Details
-
Bug
-
Status: Closed
-
Major
-
Resolution: Abandoned
-
5.8.0, 5.10.0, 5.11.0
-
None
-
None
-
Linux (Debian squeeze)
WebLogic Server 10.3.6.0 with a "Foreign Server" JMS module pointing to ActiveMQ.
WebLogic and ActiveMQ running locally.
-
Patch Available
Description
The setup:
- ActiveMQ with the default configuration providing one queue.
- WebLogic server having
- a "Foreign Server" JMS module configured pointing to the ActiveMQ's queue
- a simple EJB3 MDB hooked up to this queue via XAConnectionFactory.
What the MDB does:
It's annotated with @TransactionAttribute(TransactionAttributeType.REQUIRED), thus enforcing an XA transaction.
It calls a method from a @Local EJB, which is annotated with @TransactionAttribute(TransactionAttributeType.REQUIRES_NEW).
Depending on the JMS message content, it calls mdc.setRollbackOnly() to enforce a rollback.
What's happening in the OK case:
- you place a message to the ActiveMQ queue.
- the message is picked up by the activemq-client at WebLogic.
- the MDB is called within an XA transaction.
- when the @Local bean is called, this XA transaction is suspended (TransactionContext.end() with flags == TMSUSPEND).
- when the @Local bean returns, the transaction is continued again (TransactionContext.begin() with flags == TMRESUME).
- finally, end() with flags == TMSUCCESS and commit() is called.
What's happening in the Rollback case:
- the same until the transaction is resumed
- end() with flags == TMFAIL and rollback() is called.
But, what's going wrong is: - TransactionContext.beforeEnd() is already called during the end() / TMSUSPEND call, which in my opinion does "too much" communication with the broker, because a suspend doesn't necessarily mean that the transaction will be ok in the end.
- the message stays in the ActiveMQ queue (I have no explanation for that...).
- there's no redelivery of the message to the MDB, as if the message is somehow "lost" somewhere within the activemq-client datastructures.
- if WebLogic is shut down, then the message is really "lost" because it also vanishes from the ActiveMQ queue.
For our project, I'm going to deploy the attached quick fix, where I just will "ignore" the TMSUSPEND / TMRESUME calls to the begin() and end() methods in TransactionContext.
So the beforeEnd() method is not called during suspend and in the end every "ok" and "rollback" scenarios I tested worked perfectly for me.
I would understand if you don't take over this patch to the ActiveMQ code base, because it's in fact "ignoring" XA suspend/resume calls completely instead of implementing them properly.
However, I created this issue anyway, because the patch might be of interest for other people. Of course without any guarantees that this "works for me" patch also "works for you".
BTW:
Is a proper implementation of XA suspend/resume on any roadmap?
Thanks.