Walk me through how you would design a scalable machine learning inference pipeline that handles unpredictable workloads.
