Model Registries
A model registry is the system of record for trained models, tracking versions, stages, and metadata so deployment is controlled and auditable.
Managing models as releasable artifacts
A trained model is an artifact that must be versioned, promoted, deployed, and sometimes rolled back, much like software. A model registry is the central system that holds these artifacts and governs their lifecycle. It answers which model version is in production, how it got there, and what it was trained on, turning deployment from an ad-hoc file copy into a controlled process.
Versions and stages
Each registered model accumulates versions, one per training run promoted for consideration. Versions move through stages, commonly staging, production, and archived. A version in staging is under evaluation; promoting it to production makes it the one that serves traffic; superseded versions are archived rather than deleted so rollback stays possible. The stage of each version is explicit and auditable.
What the registry records
- The model artifact and the framework and version needed to load it
- The training run, data version, and code commit that produced it
- Evaluation metrics on a fixed benchmark for comparison
- The stage transitions, with who approved each and when
- The input and output schema, or signature, the model expects
Promotion gates
A registry is where quality gates live. Promoting a version to production can require that it beat the current production model on a benchmark, pass fairness or robustness checks, and carry a completed model card. Encoding these as gates means a model cannot silently ship worse than what it replaces, and every promotion leaves an audit trail of the evidence behind the decision.
Rollback and lineage
Because prior versions are retained with their full context, reverting to a known-good model is a stage change, not a rebuild. And because each version links to its training data and code, a production problem can be traced to its origin. The registry closes the loop with experiment tracking upstream, which produced the candidate, and with the deployment system downstream, which serves whatever the registry marks as production. See experiment tracking, dataset cards, and data versioning.