On 03/10/2015 11:44 AM, Steve Ebersole wrote:
Scott, I am not following. Pardon me if I am being dense :)
Today, in our afterCompletion hooks we have code that, in the case of
rollback, tries to make a determination of whether the rollback is due
to a timeout or a "normal" rollback. If we deem a timeout occurred,
then we delay the afterCompletion processing.
Currently, in our afterCompletion hooks, the determining code isn't
aware of the current/last application thread id. I think that we are
only tracking the first thread id that is used with the Hibernate
session but not the last thread id.
So how, specifically, would this be different in what you propose?
The new approach will not use thread id, instead in our afterCompletion
call, we will have the TM level knowledge, of whether the application
thread has called tx rollback/suspend/commit yet. If the application
thread has called tx rollback/suspend/commit already, the
afterCompletion call can safely clear managed entities from the session
without violating concurrency concerns. If the application has not yet
called tx rollback/suspend/commit yet, we need to defer the clearing of
the Hibernate session, until the TM listener calls us back, to let us
know that the application has called tx rollback/suspend/commit, at
which time we can safely clear the Hibernate session without violating
(Hibernate session) concurrency concerns.
Are you saying that with this listener approach, that the listener would
be notified of disassociation prior to the afterCompletion callback on
our RegisteredSynchronization? But in the timeout case, the
disassociation would happen later?
Yes. It is also possible that in the timeout case, that the
disassociation will come first (e.g. if there is a race between the app
committing the transaction and the reaper thread timing out) but more
likely that the disassociation comes later for the timeout case.
FYI, the TM level SPI changes  are still open for review.
On Mon, Mar 9, 2015 at 1:19 PM, Scott Marlow <smarlow(a)redhat.com
With a proposed TM level listener, we will have an SPI for notification
of when application threads associated with a JTA transaction, become
disassociated with the transaction (tm.commit/rollback/suspend time).
Having this knowledge in a synchronization callback, can determine
whether the persistence context should be cleared directly from the
Synchronization.afterCompletion(int) call or should be deferred until
the transaction is disassociated from the JTA transaction.
This idea is based on a TM level listener approach that Tom Jenkinson
 suggested. Mike Musgrove has a "proof of concept" implementation of
the suggested changes . I did some testing with  to see if the
improvement helps with clearing entities that might still be in the
persistence context after a background tx timeout.
I'm wondering if in the Hibernate ORM
Synchronization.afterCompletion(int status) implementation, in case of
tx rollback, if we could defer the clearing of the Hibernate session to
be handled by the JtaPlatform. This could be setup at
EntityManager.joinTransaction() time (if a new property like
"hibernate.transaction.defer_clear_session" is true). Perhaps via a
JtaPlatform.joinTransaction(EntityManager) registration call?
hibernate-dev mailing list