Given a distributed system, how do you handle partial failures during model training or batch inference at Google DeepMind?