If Openfire runs in a cluster, and the cluster loses one cluster node, at least some (possibly all) users cannot connect to nodes that remain in the cluster for quite a bit of time (approximately 60 seconds).
It is undesirable for a cluster to stop accepting new connections when one of the cluster nodes fail.
This occurs only when there's an unclean break from the cluster. I believe that this is unavoidable to a large degree: when running an async cluster and detecting network issues, there's a gray area where a remote node might simply be briefly unresponsive (eg: garbage collection) or actually disconnected. We probably should rely on Hazelcasts expertise here, instead of trying to tweak the params for this detection.