Monitoring & Drift

Models degrade silently. Nothing crashes, no pager fires, and six months after a successful launch the metrics have quietly sagged because the world moved and the model did not. Monitoring is what makes the difference between an ML system and an ML artifact, and interviewers treat a thin answer here ("we would retrain periodically") as the mark of a candidate who has not operated a model in production. This page covers the kinds of drift, how to detect them, the remediation toolkit, and the feedback loops that make ML monitoring stranger than ordinary systems monitoring.