PojoCache 2.0: New tree structure, optimistic locking & conc

[chasta]

Hi,
I'm experiencing a problem which I believe might be related to the new tree structure in PojoCache (i.e., the structure that enables canonical locking). It seems that when using optimistic locking, since attaching an object to the PojoCache will cause a version increment on __JBossInternal__, any concurrent transaction to one that attaches an object (even one that is seemingly unrelated and deals with completely different objects) will fail.

Consider the following code, out of a small unit test I've created for the purpose ("cache" was configured to use optimistic locking):

 UserTransaction tx = getTransaction();
 tx.begin();

 FamilyPerson father = createFamilyPerson("Rafael Gallardo", 55);
 cache.attach("/person/test1/father", father);

 Thread otherThread = new Thread () {
 public void run ()
 {
 FamilyPerson otherPerson = createFamilyPerson("Katherine Brendel", 24);
 cache.attach("/other_person/test/1", otherPerson); // auto transactional commit (optimistic locking mode)
 }
 };

 // Start the other thread and make sure it finishes before we commit
 otherThread.start();
 otherThread.join();

 tx.commit() // Fails (RollbackException)

What's your opinion? Is my suspicion correct? Is this a known issue (I didn't find it in JBoss.org JIRA), or perhaps the intended behaviour?

[ben.wang]

Hmmn... You are right this is problematic. Each time when a pojo is attached, a new cache node will be created under __JBossInternal__. So for optimistic locking this will be problem for concurrent transaction. This will evitably create high rate of rollback and degrade the performance.

I ahve created a jira:
http://jira.jboss.com/jira/browse/JBCACHE-915

A partial workaround is to use "region" in cache such that __JBossInternal__ is separate for each region. However, for concurrent access under the same region, the problem would still exist.

Any reason that you must use optimistic locking?

[manik]

Does this also affect the 1.4.x series? Specifically 1.4.1.CR1 (released earlier today) with a lot of improvements around optimistic locking concurrency? (These improvements will make it into the next 2.0.0.ALPHA2 as well)

[ben.wang]

It is not that servere in 1.4 because __JBossInternal__ is used only for shared references. So if there is creation of shared reference POJO under concurrent tx, then it will rollback. Otherwise, thing as usual.

[chasta]

Thanks for your reply.

Unfortunately, I need to use optimistic locking because in my usage scenario, deadlock prevention and/or recovery will be very hard when using pessimistic locking... That was one of the reasons for us to use JBossCache in the first place...

Any chance this will get fixed in ALPHA2 ?