What it is
A model registry is a catalog of trained models. Each entry records a model version along with its metadata, metrics, and the stage it currently lives in. It is the system of record that answers which model is in production right now.
What it stores
For every model the registry usually keeps:
- A unique name and version number
- A link to the trained artifact such as the weights file
- The metrics and dataset used to evaluate it
- A stage label such as staging, production, or archived
Stage transitions
A model is promoted through stages as it earns trust. A new version starts in staging where it is tested, then it is promoted to production when it passes, and the old version is archived. These transitions are logged so you can audit who promoted what and when.
Why teams use it
- It makes rollbacks easy because the previous production version is one click away
- It separates training code from deployment, since serving just asks for the production version
- It gives a clear history for governance and reproducibility
Key idea
A model registry is the versioned source of truth for which model is in each stage and how to roll back.