Transactions problem

[qwier]

Hi all,
I am trying to figure out how to get transactions work with the cache.

Please have a look at the following code snippet:

InitialContext ic = new InitialContext();
UserTransaction ut = (UserTransaction)ic.lookup("UserTransaction");
ut.begin();
Integer x = (Integer)cache.get("/Test","TestInteger");
x = x+1;
Thread.sleep(5000);
cache.put("/Test", "TestInteger", x);
ut.commit();

This code is simultaneously running from two different JBoss instances within one cluster.
How can I manage to get the x value increased by 2? Now it is increased by 1.

Below are my cache settings related to transactions:
TransactionManagerLookupClass=org.jboss.cache.DummyTransactionManagerLookup
IsolationLevel=SERIALIZABLE (actually here I tried all of the possible values)
CacheMode=REPL_SYNC
NodeLockingScheme=PESSIMISTIC

Thanks!

[manik]

If they run simultaneously, both transactions won't succeed. One of them will fail since there will be a race condition to commit at the same time, and both transactions will have locks on the node in question.

If you removed your sleep time, chances of success are higher, but there will still be that race condition. This is expected since cluster-wide locks are not obtained when you read your node using a cache.get().

[qwier]

If you removed your sleep time, chances of success are higher, but there will still be that race condition. This is expected since cluster-wide locks are not obtained when you read your node using a cache.get().

Thanks!
But how can I obtain cluster-wide locks when modifying data inside the cache?

[manik]

You can't. We don't do cluster-wide locks. :-) Remote locks are obtained on transaction commit time.

[qwier]

You can't. We don't do cluster-wide locks. :-) Remote locks are obtained on transaction commit time.

OK., Lets assume a situation where N services on N different JBoss instances are reading and modifying data, how data integrity can be achieved? Only by catching failed transactions and re-running them again?

[manik]

Yes.

You would want SYNC_REPL though.