...
Maintainability: decoupling the cluster code from the broker and assuming the broker behaves deterministically makes it very easy for developers working on the stand-alone broker to unintentionally break the cluster, for example by adding a feature that depends on timers. This has been the case in particular for management, since the initial cluster code assumed only the queue & exchange state needed to be replicated, whereas in fact all the management state must also be replicated and periodic management actions must be co-ordinated.
Non-replicated state: The current design replicates all state. In some cases however, queues are intended only for directly connected clients, for example management queues, the failover-exchange queues. It would be good to be able to define replicated and non-replicated queues and exchanges in these cases.
...