XAException.XAER_RMERR should be treated as an error

Yes, that's a bug in the 1PC optimization's error handling logic. I'll fix this ASAP.

I've just pushed the changes to fix this bug and published a new 2.1.3-SNAPSHOT version in codehaus' snapshots repo.

Could you please have a look at it and let me know if this works up to your expecations? As a reminder, the instructions to access codehaus' snapshots repo is at http://docs.codehaus.org/display/BTM/Maven2#Maven2-Earlyversions

It works well, the TransactionManager.commit() throws a RollbackException which the caller can catch. And additionally it can extract the wrapped SQLException's information from it.

But there is an other problem: the transaction manager's rollback() method throws an IllegalStateException with a message "no transaction started on this thread" after the commit() throws a RollbackException so the failed transaction cannot be rolled back properly (and it will not rolled back internally). If this is a separate issue then I will create an new issue about it.

I don't think there's any other problem, at least if I understand you right.

TransactionManager.commit() terminates the transaction, no matter what. Its goal is to commit it but if it fails to, it can throw RollbackException after cleaning up the transaction.

In a nutshell, there's no need to call TransactionManager.rollback() when TransactionManager.commit() has thrown an exception: the commit method already did the job of the rollback method and that why it threw RollbackException: to indicate this fact.

This is all strictly compliant with the JTA spec.

It's alright when the commit() fails with a RollbackException then it rolls back the actual transaction, but the problem is that it does not do this. The underlying database transaction remains uncommitted/waits for to be rolled back.

If the underlying database transaction really remains uncommitted, then there is a bug in your database and/or JDBC driver.

Quoting http://docs.oracle.com/javaee/1.4/api/javax/transaction/xa/XAResource.html#commit%28javax.transaction.xa.Xid,%20boolean%29 :

If the resource manager did not commit the transaction and the parameter onePhase is set to true, the resource manager may throw one of the XA_RB* exceptions. Upon return, the resource manager has rolled back the branch's work and has released all held resources.

What version of Postgres and its JDBC driver are you using? With that information I'll be able to investigate the driver's source code to try to figure out if that's a geniune bug or a mis-understanding of the JTA spec.

I suggest you to report this problem to the postgres JDBC developers and copy me on the conversation so I'll be able to follow on what's going on and eventually argue if that's needed.

You are right, I did not read the JTA documentation properly.

I use the most recent stable version of the PostgreSQL JDBC driver (9.1-901). It's XAResource implementation is the org.postgresql.xa.PGXAConnection class. It contains a commitOnePhase() method which is called when the commit()'s onePhase parameter is true.
And it does not roll back the transaction when a SQLException occurs.

I dig around a little if it's really a bug and report it to the PostgreSQL JDBC driver developers.

I've created an issue about it. If you want to you can follow it here: https://github.com/pgjdbc/pgjdbc/issues/2