The only reason we don't want both RMs to be active is to avoid the split-brain situation. Today, if both RMs become active, one of them fails to create the fencing-node and should automatically transition to standby.
Actually, this behavior could confuse user as it sounds like which RM is chosen by randomly. The expected behavior here is: if user doesn't specify "--forceactive" which means user may not know other RM get activated now - nothing should get updated by YARN and a warning message from CLI get thrown out if already another RM in activated state. No?
I wonder if this affects the latency of transitioning active by any significant amount.
The only latency to add here is reading a YARN configuration which doesn't sounds like a problem for user experience especially we support RM down with work preserving.