Even within a replica set, we use the ReplicaSetMonitor to target nodes based on read preference for internal clients (like the LogicalSessionCacheRefresher). While the vast majority of times this will be correct and up to date with the ReplicationCoordinator's topology, there may be times when it is behind or (due to a bug) is out of sync entirely. This can lead to unavailability if the RSM's view of topology is incorrect.

We could consider using the ReplicationCoordinator as the source of truth on replica sets, or an least periodically checking and invarianting if the ReplicationCoordinator state is out of sync with the RSM topology.