Very glad to see percentage-of-users experiments, that's where it falls short to simply deploy/rollback the new version.