Description
The scenario is as follows:
- message is delivered within an XA global transaction
- the XA tx completes the prepare phase
- the server crashes before the XA tx completes the commit
- the server is restarted
- (if the J2EE server supports XA recovery, it will recover the XA tx and complete the commit (tested and confirmed with Websphere))
- ActiveMQ redelivers the message
The redelivery should not happen. Unless, following recovery, the TransactionManager decides to rollback the tx, the tx will be committed. If it is committed, the message would have been processed already. Redelivering it causes the message to be processed again and breaks JMS' deliver-once-and-only-once guarantee.
If the tx has been prepared, AMQ should not redeliver the message following a crash recovery, If the tx is ultimately rolled back and the message is put back on the queue, it can redeliver it, otherwise, it should not.
This is very difficult to write a Junit test for but I will try to come up with something...
Attachments
Attachments
Issue Links
- is duplicated by
-
AMQ-3305 Implement "exactly once" delivery with kahaDB and XA in the event of a failure post prepare
- Resolved