What a registry is
A model registry is a central store of trained models with versions and lifecycle stages. Where experiment tracking records how a model was made, the registry governs which version is allowed to serve.
What it tracks
- Versions of each named model, incremented on every registration.
- Stages such as staging, production, and archived.
- Lineage linking a version back to its training run and data.
- Approvals recording who promoted a version and when.
Why teams need it
Serving systems should pull the model by name and stage, not by a file path. Promotion from staging to production becomes a controlled, auditable action rather than someone copying a file. Rollback is just demoting one version and promoting the previous one.
The registry is the single source of truth that connects training, approval, and serving.
Key idea
A model registry is a versioned, staged catalog that controls which model serves, making promotion and rollback auditable rather than ad hoc file copies.