When one of our GPU providers performed maintenance, we asked to migrate traffic to a different cluster beforehand to avoid disruptions. During the migration back, the provider encountered an issue that led to some conversations experiencing an elevated startup time.