When several models, versions and experiments exist, you need to know exactly which model is running and how it was made.
What a Model Registry Stores
- Model artefacts: the trained files.
- Version numbers for each trained model.
- Metadata: training date, owner, framework and dependencies.
- Lineage: code version, training data version and hyperparameters.
- Evaluation results, including performance by group.
- Stage: for example development, staging, production or archived.
- Approvals and change history.
Why It Matters
- Reproducibility: retrain or inspect any version.
- Rollback: switch back to a known good model quickly.
- Auditing: show which model made a decision and how it was validated.
- Collaboration: a single source of truth for teams.
Promotion Workflow
- Train and register a candidate.
- Evaluate against the current production model on the same test data.
- Review — automated checks and, for important models, human approval.
- Promote to production, keeping the previous version available.
Tools
Registries are built into experiment-tracking tools (such as MLflow) and cloud ML platforms. Small teams can start with disciplined naming, object storage and a metadata file, then adopt a registry as needs grow.
Link Everything
Record the model version with every prediction logged in production, so any decision can be traced to the exact model that made it.