← AI Terminology
Model Registry
A model registry is a central catalogue of trained model versions with metadata, lineage, stage labels (staging/prod), and approval workflows.
It is the source of truth for what runs in production.
It is the source of truth for what runs in production.
Why It Matters in AI
Without a registry, teams lose track of which checkpoint is live. Registries enable audit, rollback, and promotion gates — mandatory governance for regulated ML and helpful for all AI products.
Key Points
| Aspect | Description |
|---|---|
| Tools | MLflow Model Registry, W&B Artifacts, cloud registries |
| Access | RBAC on who can promote to prod |
| Stages | None → Staging → Production → Archived |
| Stores | Weights refs, metrics, signatures, lineage |
| Related | MLOps, experiment tracking |
| LLM twist | Also track prompts, LoRA adapters, base model IDs |
Simple Analogy
A library card catalogue for models — every edition labeled, who approved it, and which branch is checked out to the reading room (prod).
Common Usage Examples
mlflow.register_model- Promote v12 to Production stage
- Deploy pulls only Production alias
- Lineage: dataset v3 → model v12
Summary
In short: A model registry catalogues versioned models and their stages — the control plane for safe promotion and rollback.