ระยะ: Staging, Production, Archived

A version number alone does not say which model is live. Stages add a label that tells everyone where each version sits in its lifecycle. 🚦

MLflow gives each version one of four stages: None, Staging, Production, or Archived. A fresh version starts at None until you move it.

Staging is your dress rehearsal. A model here is being validated and shadow-tested, but real users are not relying on it yet.

Production marks the version actually serving users. Your inference code loads this stage, so promoting here changes what people really get.

Archived keeps an old version around without serving it. You preserve history for audits while clearly signaling it should not be used.

Use the client to transition a version into a stage. One call moves version 3 of your model into Staging.

When you promote to Production, pass archive_existing_versions so the previous live version is retired in the same move, with no gap.

Keep exactly one version in Production at a time. That way loading by stage is unambiguous and nobody guesses which model is serving.

Newer MLflow favors aliases like champion over fixed stages. An alias is a movable nickname you point at any version you choose.

If a promoted model misbehaves, you just transition the previous version back to Production. Rollback is a label change, not a redeploy.

Treat each transition as a gate. Require passing tests or approval before a version is allowed to reach Staging or Production.

Pick the stage your serving code should load.

Stages turn a pile of versions into a clear lifecycle: None to Staging to Production to Archived. Promotion and rollback are now just safe label changes. ✅