Imagine the following scenario with a sharded cluster with 1 router, and 1 shard with 3 nodes and hedged reads active:

• An insert occurs and it's majority written, but haven't been replicated on one of the nodes
• A count operation is performed
• The fastest request is the one to the node that haven't replicate the insert yet
• A StaleDB or StaleShardVersion error is thrown

The current code only ignores MaxTimeMSExpired errors so a stale error will make the operation fail.