I have more information from my tests.
If I shutdown the first server, the error stops happening on the second one.
The conclusion I can get is that the MDBs from both servers are acessing the queue on the first loaded node - what is correct. As the JMS service is HA singleton.
With this info, the error could be fixed by changing the <mdb-user> and/or <mdb-subscription-id> from the MDBs on the xmls of the second server.
But - it does not sound nice to have to generate different jars/ears for each node of the cluster. The farm service for example would not be usable.
A durable topic subscription and/or client id cannot be used by two clients at the same
If you want to share the messages then use a queue.
If you just want HA for the MDB that processes the topic subscription
deploy your mdb in deploy-hasingleton so it only runs on one server.