Model Registry — AI Glossary

Reference details

Related terms

• Model Hub
• Experiment Tracking
• MLflow

How it works

1. Registering an artifact mints an immutable version number under a model name, so a version always refers to one specific set of bytes and is never reused.
2. The registry stores a reference to the artifact rather than a copy of it, which is why an entry can survive its underlying file being deleted and prove nothing.
3. Metadata is attached at registration: the producing run, the dataset snapshot, the evaluation result, and the approving actor — captured then because none of it is recoverable later.
4. Stage transitions are recorded as events with an actor and a timestamp rather than as a mutable field, so the promotion history is itself the audit trail.
5. Deployment reads an alias rather than a version number, which is the indirection that makes promotion and rollback a pointer change instead of a rebuild.
6. A rollback is therefore repointing the alias at the previous version, and it only works if that version's artifact still exists — which is what retention policy on the artifact store actually governs.

Example: The rollback test

The question a registry has to answer is: production is behaving badly, what exactly is deployed, and how do I get back? A useful entry pins the artifact hash, the training or fine-tuning run that produced it, the dataset snapshot, the evaluation run that cleared it, and who approved the promotion. Rollback then becomes repointing the production alias at the previous version — seconds, no rebuild. Without it, rollback means reconstructing a build from memory during an incident, which is when reconstruction is least reliable.

Commonly confused with

Three things get conflated. Experiment tracking records runs and metrics during development. A model registry governs which versioned artifact is promoted to which stage. A model hub distributes models to other people. MLflow ships the first two together, which is why the distinction blurs.

When to use it

Reach for it when:

• More than one model version can be in production at once, or you run staged rollouts
• Promotion needs an approval trail — regulated settings, or anything customer-facing
• Fine-tuned models, where the artifact is yours and nobody else can re-serve it for you

Reach for something else when:

• A single hosted third-party model behind an API — pin the model string in config instead
• Prototypes, where the overhead outweighs the traceability
• As the only lineage record: an entry pointing at a deleted dataset proves nothing

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Model Registry — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path