When mongos cannot connect to any of the members of a shard replica set for extended periods of time (> 5 minutes), it will remove the ReplicaSetMonitor for that particular set in memory. The consequence is that it will start returning "unknown replica set" error instead of the usual cannot connect to host X.

Original summary:
'unknown replica set' error message from mongos during failover
Original description:

We had a user report that he got an error from mongos during a failover:

unknown replica set d303f52f76534680956e00a707bbca89 (71)

This doesn't seem to be an error that should make its way back to the user. Is this possibly a bug in mongos?

Here is the relevant report from the user.