This is a list of known issues (mostly surrounding the _check method) that should potentially be tackled at the same time.

• When selectAndCheckNode or getMaster can't find a suitable node, it calls _check to update the local view of the set which in turn calls _checkConnection and waits on _checkConnectionLock. This means that only one thread at a time can be issuing isMaster calls. Each thread will wait its turn to issue an isMaster call to each node in the set even if another thread has already updated while we were waiting. This can result in significant latency as a request can block for thousands of threads to do up to 24 round-trips over the network (2 per node in the set).
• _check unconditionally retries after sleeping for 1 second if no master is found.
  • selectAndCheckNode has to wait here even if there is a suitable secondary that it could return.
  • This can delay the ReplicaSetMonitor background thread (which calls checkAll -> check -> _check) preventing it from checking other sets, even though it will retry with that set again soon.
  • The sleep also happens after the retry even though it won't retry again.
• _check iterates over _nodes but indirectly calls _checkHosts which can mutate _nodes, causing hosts to be skipped. There is some protection against this from _checkConnMatch_inlock, but it isn't complete.
  • There is a related issue that it will alternate between adding and removing a host if two nodes disagree about who is in the set. This might cause an issue in a split-brain scenario where one reachable host doesn't think the current master is in the set (especially since _master is just an index into _nodes).
• Updates due to needing a master should probably take advantage of the "primary" field in isMaster responses to find the master rather than linearly scanning _nodes.

Not bugs, but changes that could make the class easier to reason about:

• Both of the getSlave methods are unused and can be deleted.
• If getMaster() just became a wrapper around selectAndCheckNode(PrimaryOnly) all node selection and retry logic would be in a single path.