How would you design a robust ML pipeline for production, considering data drift and model monitoring?