[
https://issues.apache.org/jira/browse/OPENJPA-343?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12524717
]
Kevin Sutter commented on OPENJPA-343:
--------------------------------------
> Here's where we might disagree. The user-level commit should fail so the user
> doesn't think everything is ok as far as the cache is concerned. I understand
> that the database transaction is complete and whatever changes have been made
> there are permanent. But the EntityManager is possibly corrupted.
I agree. We are performing this "afterCompletion" processing as far as we can
until we hit this unexpected exception. This exception is now being logged (if
trace is turned on) and the exception is returned to the caller. Unless you
are suggesting that we should possibly log-and-eat this exception and attempt
to continue additional processing as if nothing has happened, I think we have
done everything we can do. If we went this log-and-eat route, it would require
more more granular try-catch blocks in this path. Not sure this is necessary
processing for the unexpected (rare) case.
FYI, we are still flowing through the EntityManager with the setRollbackOnly
invocation. So, any processing in the EntityManager and/or Broker that would
be triggered because of the setRollbackOnly call will still happen. It's just
the explicit setRollbackOnly call on the Transaction object itself that was
conditionally skipped if the Transaction was not "active". Just wanted to
clarify that aspect of the change.
If there are still concerns about the general exception processing in the
OpenJPA code base, then maybe we should open a separate JIRA issue or at least
start a [DISCUSSION] topic on our dev mailing list. The original intent of
this Issue was to not call the setRollbackOnly method when the Transaction is
not in a state to accept the invocation. I think that issue has been resolved.
But, maybe there are still concerns about the general exception processing
within OpenJPA...
Thanks,
Kevin
> Do not call setRollbackOnly on inactive Transactions
> ----------------------------------------------------
>
> Key: OPENJPA-343
> URL: https://issues.apache.org/jira/browse/OPENJPA-343
> Project: OpenJPA
> Issue Type: Bug
> Components: kernel
> Affects Versions: 0.9.7, 1.0.0
> Reporter: Kevin Sutter
> Assignee: Kevin Sutter
>
> While in the middle of processing an afterCompletion invocation in
> BrokerImpl, an unexpected RuntimeException (IndexOutOfBoundsException)
> occurred within some underlying WebSphere code. While we (OpenJPA) were
> attempting to clean up after that exception, we attempted to call
> setRollbackOnly on the current transaction. But, since we were in the
> process of completing the current transaction, it is invalid to be calling
> setRollbackOnly and we ended up getting an IllegalStateException from the
> WebSphere Transaction Manager. Due this second exception, we ended up losing
> track of the original exception and this became a difficult problem to
> diagnose.
> This issue will be used to correct a couple of issues (at least):
> 1) We should ensure that the transaction is active before calling
> setRollbackOnly(). When an exception happens during afterCompletion
> processing, the Transaction can no longer accept setRollbackOnly
> invocations.
> 2) When an unexpected exception happens like this, we should log the
> exception before attempting to process the exception. In this particular
> case, we lost the original exception when we ran into the
> IllegalStateException
> from the Transaction service. This forced us to re-run the scenario just to
> get a trace of the exception.
> 3) Or, if we don't want to log the exception immediately, we need to
> determine why we lost the first exception in the first place and ensure that
> doesn't happen again.
> Kevin
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
