← Back to concept explainers

Concept Fable

Model registry

Conclusion first:A model registry is a system for centrally managing model versions, metadata, approval status, and deployment flow. It usually records model artifacts, versions, training sources, evaluation results, owners, risk levels, stage status, and deployment targets. For enterprise AI platforms, a model registry is the control plane for models from experimentation to production, rollback, retirement, and auditing.

Model Registry · MLOps / platform

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

Unregistered boats cannot leave port

Tidegate Harbor had many shipwrights. Some built fast boats, some built cargo ships, and some refitted old vessels. In the early years, the harbor was small; once a boat was built, it could be launched, and the dockmaster let it pass based on familiarity.

Later the trade routes multiplied, and so did the ships. Several ships were named Petrel—some could sail far, some only near shore; some had already had their keels repaired, some were still on sea trials. The dockmaster began to lose track of which ship could take which kind of cargo.

The port's first fix was to hang colored flags on the bow. A red flag meant a new ship, and a blue flag meant an old ship. But flags could be blown off by the wind, and some people temporarily switched flags to rush orders.

After a storm, the port investigated which ship had carried valuable cargo without inspection. The ledgers, dock records, insurance policies, and pier release stamps each told their own story, and no one could piece together the full lifecycle.

The new port official established a ship registry. Every vessel entering the registry had to record its structure, build batch, sea trial results, responsible person, risk level, permitted routes, approval status, and decommissioning date.

The ship registry office does not build ships, but it determines the state transitions of a ship from trial, candidate, available, production, paused, to retired. Anyone who wants to send a ship on a new route must check the registration status.

Afterwards, the shipwright can still iterate quickly. Different versions of the ship can coexist, but every promotion, rollback, freeze and replacement is recorded, and the dock no longer clears ships based on memory. When a customer complains, it is also possible to trace which ship it was, who approved it, and why it was allowed to leave port.

The harbor master finally said that what the port truly needed was not more ship names, but a trusted place that told everyone which ship was in what state and whether it could be assigned to a real route. The ship registry also kept decommissioning records of old ships. Not out of nostalgia, but so that when something went wrong, people could know which version of a ship had once served a route, when it was replaced, and why. Without that timeline, the port could only piece together the truth from memory after an accident. Later, the dock chief no longer cleared ships just because they looked familiar. Ship names can repeat, rumors can be wrong; only registration status, the responsible person, and sea trial results can determine whether a ship should carry cargo. Later, during the retrospective, everyone wrote this lesson into daily rules: first look at where on site the result actually changed, then decide what to keep, what to repair, and what to restrict. If you only look at surface excitement, you will fall in the same place next time.

Reveal

This story is about: model registry

A model registry is a system for centrally managing model versions, metadata, approval status, and deployment flow. It usually records model artifacts, versions, training sources, evaluation results, owners, risk levels, stage status, and deployment targets. For enterprise AI platforms, a model registry is the control plane for models from experimentation to production, rollback, retirement, and auditing.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Ship registry: model registry
  • Several ships are all named Petrel: chaotic model naming and unclear versions
  • Color flags: unreliable temporary status markers
  • State transitions from trial to retirement: lifecycle stages such as staging, production, archived
  • Permitted routes: model scope of application, deployment environment, and risk boundaries
  • Promotion and rollback records: promotion, rollback, and approval trail
  • Complaint investigation: auditing, traceability and accountability
  • Final insight: a model registry provides a trusted control plane for the model lifecycle

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

Model versioningartifact storeMLflowlineagepromotionrollback
京ICP备2026014272号