In a 3-node cluster hosted on OpehShift 4.7, messages stop redistributing, even though there is no matching local consumer on the node where messages are parked and there is a matching consumer on a remote node. The strange thing is that some messages are redistributed and delivered to the end client, but some are left parked. Looking into the heap dump for the node where messages are parked, we can see that the redistributor for the matching queue is non-null. No selectors / filters are in use by the client. This has been reproduce with the artemis CLI client using a topic producer and a queue consumer with FQQN. All addresses and queues are static and auto-creation is disabled.

The interesting thing is that new messages do seem to get distributed to the end consumer and some old messages (pre-dating the attachment of the consumer) also get consumed, but a few messages remain. This pattern varies from incident to incident - sometimes no old messages are redistributed, sometimes some are redistributed and some are left.