The test started to fail due to the change of the test code - it used to
NOT
close the entity manager before triggering the fetch group.
That means
it's not a regression - to fix the test it should be just rolled
back to its
previous state (rev. 6213).
The test was altered to expose the bug, which
happened when a fetch group is
triggered on a detached object.
That bug
was fixed as part of work on nested fetch groups (see
bug
244124).
However this is still happening on
WebSphere.
Investigation showed that on WebSphere after the entity
manager is closed the
object is still registered in the entity manager's
unitOfWork (and the uow is
not dead).
That is caused by fix to
Bug
259993 - em.find() hangs up in some situations on
WebSphere
7.0.0.1.
WebSphere apparently caches EntityManagerImpls - instead of
closing them it
calls em.clear and then return to a cache for future
reuse.
The problem was the timing of the clear call - it came when
Eclipselink has not
yet merged uow (in afterComplete).
Fix for
Bug
259993 cured this problem but not allowing em.clear() called
while
uow is merging to do anything but to clear the cache.
There are
several drawbacks to this solution:
1) Objects read in by ValueHolders
triggered on detached objects may loose
identity (uow has no cache);
2)
uow is never released - map of clones, etc is held until the uow is
garbage
collected.
This second point is also the reason for the test
failure - the fetch group
code verifies if the object is registered in the
uow, and if it is assumes that
the object is also present in uow's
cache.
But currently on WebSphere using jta after em.close() the object is
still
registered, but is not in the uow's cache.
The suggested
solution: if em.clear is called on synchronized uow, and if the
uow is
merging (WebSphere's case) then don't clear the cache, but set the flag
that
will cause uow to be released after the merge in afterCompletion
is
complete.
That would cure both problems listed
above.