Thanks for the input and clarification, Justin! I found a misconfiguration of the property redistribution-delay in the cluster, that's why it did not behave the way I expected. Now it behaves as expected, and as you explained.
Regarding the reason why we are implementing a cluster: We are not coming from ActiveMQ 5.x, and we have not yet done any benchmarks of using a cluster. The reason for evaluating how a cluster is setup and performs is because the future load on the broker is unkown to us at this time. So we want to be prepared for any type of scenario coming up. But it might be that we run a single broker in the end (at least to start with). Thanks for the help! best regards Magnus