Currently cluster connections use discovery of one form to create a cluster. Once created a cluster connection will reconnect when a broker goes down and the default attempts for this is -1 which is forever. In an environment where the brokers are ephemeral and the cluster size changes to handle scale this doesn't work well. You can set up say 10 reconnect attempts, which t is the default in Openshift, but this is arbitrary and if a broker doesn't come back up in time then you have the opposite problem where connections are torn down then recreated differently which can have adverse effects.

The real fix here is for the endpoints to be explicitly defined, this way the operator can just update the list and the broker can react accordingly.