mod_cluster maintains affinity (aka session stickiness) between clients and application servers. However, if scaling up or down number of http / frontend servers running mod_cluster, a client may be directed to 1 server intitially and then be directed to another frontend later (depending on connection distribution in front of http servers). If this occurs, other http servers are not aware of each other's affinity, and so there is a chance that (if session is not replicated on jboss as servers) a clients session will be lost if he is migrated to a new http server. This would be particularly applicable in a cluster consisting of a number of buddy partitions fronted by >1 mod_cluster http servers.

Not sure about implmentation, but if the mod_cluster.sar service could report relevant info required to maintain affinity back to all frontend servers, a higher level of fault tolerance could be achieved.