When a mongod or mongos is being hit with a huge number of connections, it may run into the operating system's limit (configured or other) for creating new threads, and thread creation may fail. For example,

ERROR: Uncaught std::exception: boost::thread_resource_error, terminating
dbexit:  rc:100

Rather than letting this be an "uncaught" exception, we could catch it, log an explanatory message, close the new connection and stay running. This would be more desirable behavior from the customer's perspective, and should be easy to test during development and in smoke tests.