1. I have some doubt here. Since non-persistent messages are not written to disk, failing over to the passive broker should lose any messages still in the broker's memory. Does the test use a slow consumer to make sure there is a backlog of messages on the broker at the time of failover? Note the producer, via the failover transport, should resend any message for which it did not receive the broker's response (I want to say it will, but the logic there is beyond my understanding).
BTW - are the Topic subscriptions in the test durable? It shouldn't matter, but I'm curious. 2. Not sure on performance. One thing I do know - if there are a lot of messages in the persistent store, KahaDB can get extremely slow at startup - for some reason, it seems to read through the entire store of messages at broker startup. We've seen 20 minutes waiting to read through about 10gb of store, and longer with larger DBs. I believe LevelDB is supposed to be better in that regard. 3. If the lock-check-delay is set too low, delays in network traffic can trigger dual active brokers - a very, very bad scenario as they will trash the kahadb and step all over each others messages there. Even 10 seconds could be too short in some situations. -- View this message in context: http://activemq.2283324.n4.nabble.com/Failover-and-non-persistent-messages-tp4676862p4676916.html Sent from the ActiveMQ - User mailing list archive at Nabble.com.