Model Registry
A versioned catalog of trained models — their versions, metrics, and lineage — the source of truth for what is deployed.
A model registry is a central, versioned store for trained ML models and their metadata — version, training data and code, evaluation metrics, and stage (staging/production). It is the source of truth for which model is where, and the control point for promoting and rolling back models (MLflow, SageMaker Model Registry).
Worked example: a new model is registered as v5 with its metrics; a reviewer promotes it from “staging” to “production,” triggering deployment — and if it misbehaves, you roll back to v4 by flipping the stage, with full lineage of what changed. Gotcha: a model is not reproducible from weights alone — you must version the training data and code alongside it, or you cannot explain or recreate a model later; the registry is where that lineage lives, turning “which model made this prediction?” from a mystery into a lookup.