The problem was that the timeout for get_primary() is 30s and that the connection timeout for mongo_client is also default to 30s for the test. So if it blocks on connecting to a node, the loop of finding the primary is also stuck. And so it times out first before being able to check isMaster on all nodes. I think we should set a smaller connection timeout (timeout_millis) when looping through all the nodes so we don't wait on a bad node forever.