The main reason is because allowing these exceptions to be (re-)thrown
opens up mixed heuristic issues.
Also bear in mind that we have no clue about the nature of the process
being performed by the synch. It highly likely that "failure" of the
synch should *not* rollback the main transaction. This is really one of
the 50/50 data points. We really just do not know, and the contract
gives us no way to know. However, if rolling back is that important it
is in fact possible in any case i can imagine to pass the
o.h.Transaction into the Synchronization constructor and actually roll
it back if an error occurs.
On 03/24/2010 03:54 PM, Adam Warski wrote:
Hello,
if a transaction synchronization throws an exception, is it only logged, and not thrown
further (see org.hibernate.transaction.JDBCTransaction, line 273). Is there some reason
for this?
As Envers uses tx synchronizations quite extensively, when an exception is thrown in the
synchronization, I roll back the transaction manually. So, no data is persisted (which is
the desired behavior), but the client isn't notified in any way that something went
wrong; for the client, the operation behaves as if the tx commited successfully.
I suspect that maybe some applications rely on the fact that the exception is eaten and
not re-thrown. If there are no contra-arguments to throw the exceptions, maybe a good
solution would be to re-throw the exception is the transaction is already marked for
rollback? Or if it was marked for rollback in the synchronization?
The related JIRA issues are:
http://opensource.atlassian.com/projects/hibernate/browse/HHH-3543
http://opensource.atlassian.com/projects/hibernate/browse/HHH-4721
By the way, how does Hibernate Search deal with such situations? I looked at
PostTransactionWorkQueueSynchronization, and it seems that it's possible that the
transaction commits, but the data isn't indexed properly, if the
queueingProcessor.performWorks throws an exception?