If a network partition happens between app servers and the primary of a replica set, no failover will occur despite the fact that secondaries are reachable. There may be useful ways of detecting this by sending the client connectivity state to all members of the replica set at intervals, for example.