Repro:

1. Set up two shards, each with one primary, one secondary, and one arbiter.
2. Enable sharding on a database.
3. Shard a collection in the database.
4. Insert enough documents to distribute them between the shards.
5. Run a test that constantly runs various queries against the collection.
6. While running the test kill a random mongod then restart it.

After killing the mongod you may or may not see errors like this:

database error: DBClientBase::findOne: transport error: ...
database error: socket exception
database error: dbclient error communicating with server: ...

These errors are passed back to the client application by mongos. Is this the expected behavior?

This was found while trying to reproduce PYTHON-212.