A model registry sounds like a fancy word for a spreadsheet, and for the first six months at most companies, that’s exactly what it is. A shared folder full of model_v2_final_ACTUAL.pkl files, a Slack thread deciding which one is live, and one engineer who happens to remember which checkpoint got rolled back in March. It works until the day it doesn’t — usually right after someone promotes the wrong artifact to production at 2 a.m. and spends four hours figuring out which of nine near-identical .pt files is the one currently serving traffic.
(Knowledge cutoff note: registry tooling in this space — MLflow, Vertex AI Model Registry, SageMaker Model Registry — moves fast. Version numbers and feature sets below reflect general architecture, not a specific release; check current docs before treating any tool-specific detail as current.)
What a Registry Actually Does
A model registry is a system of record for trained models: it tracks every version of every model, the metadata that describes it (framework, training data hash, evaluation metrics, who trained it and when), and — the part people underestimate — its current lifecycle stage. Stages are typically something like None → Staging → Production → Archived, and the registry enforces transitions between them rather than just labeling them.
