We should start treating warnings as errors in our Python tests, and collect a list of warnings we don't care about. I'm preparing a PR for this that is catching quite a few bugs, both in datalab and in our deps.