Atomikos Forum

I'm using atomikos as JTA implementation on Tomcat and Oracle as Database. When I open an XA Transaction enlisting 2 or more datasources and performing only readonly operations (selects and/or updates/deletes that don't affect any rows) atomikos leaks all connections I used in this transaction. They are never released or reaped back to be available in the ConnectionPool.
Debugging atomikos code I noticed that it happens because in the ActiveStateHandler.commit() method, when the call to prepare() returns Participant.READ_ONLY, it doesn't call commitWithAfterCompletionNotification(). I understand that it makes sense not running the second phase if the first says that it's a readonly transaction, but it has two side effects:
1. AtomikosConnectionProxy.afterCompletion() is never called and thus the connection is not released.
2. This transaction state is never changed from BranchEndedStateHandler to TerminatedStateHandler, which makes the reap method to not release this connection during maintenance as well.

Just one more information, other unexpected behavior I noticed in this case is that although the ConnectionPool.reap() method is unable to release these connections it still updates the lastReleased timestamp, what makes it look like the connections are still in use if you look at the debug logs when the maintenance runs.

Hi,

I discovered the same problem today. You can solve it temporary by creating two separate transactions (one for each datasource). Not the best solution but working for me.

Yes. If you have only one participant in a transaction atomikos always do one phase commit and there's issue in this case.
We fixed the issue (by notifying all synchronizations after a readonly transaction) and rebuilt transactions.jar. Now I'm waiting for a reply from Atomikos team in order to make the fix available to the community.

Hi,

This sounds like a likely bug indeed. I created a case for this.

Thanks!

Hello Guy,

What's the case ticket number? Is it accessible publicly?

The case no is 120952 - but not accessible publicly (alas).

Hi,

I added a test in order to reproduce this bug, but failed. The Synchronizations are notified correctly - at least in the latest 3.8 and 3.9 support branches.

What release are you using?

Guy

I'm using Atomikos Essentials 3.9.3 and Oracle ojdbc6 driver (11.2.0.4). It's pretty easy to reproduce the issue in this version. I can send you a test code if you want.

The only thing you need to make sure is that at the end of ActiveStateHandler.prepare() method, where you have the following test, allReadOnly is true:
        // here we are if all yes.
        if ( allReadOnly ) {

Then it's going to change the handler to TerminatedStateHandler  and return. The method that called it, (CoordinatorImp.terminate) is going to get Participant.READ_ONLY as response and not to call commit().

Hi,

I added a test with two readonly XAResource instances and failed to reproduce.

So yes, if you can share a test with us that would be great. Just send by email to support at atomikos dot com and use "case 120952" as the subject line.

Thanks!