Model drift detection

Design a model-monitoring and drift-detection system for an organization running hundreds of production models, where labels for many models arrive late or never. It must detect feature drift, prediction drift, and performance degradation, distinguish a real model problem from an upstream data outage, and alert the owning team with enough context to act, all without adding meaningful latency to the serving path. Cover what you log, the detection methods, and how you avoid alert fatigue.