Registry Automation & Promotion Gates

The registry as a control plane

Manually clicking 'Promote to Production' doesn't scale and leaves no audit trail. Drive promotion from code so every transition is reviewed, logged, and reversible:

from mlflow import MlflowClient
client = MlflowClient()

# Promote only after the gate passed in CI
client.set_registered_model_alias(
    name='credit-scoring', alias='champion', version='8',
)
client.set_model_version_tag(
    name='credit-scoring', version='8',
    key='promoted_by', value='ci-pipeline-run-4821',
)

Wire it together: CI trains → mlflow.evaluate gate passes → pipeline moves the @champion alias → serving picks up the new version on its next poll. A failed gate simply leaves the alias untouched. Rollback is one alias re-point, fully audited via tags.

Alias promotion is a railway signal thrown by an interlocking, not a person standing on the track. The interlocking (CI gate) refuses to clear the signal unless the line ahead is proven safe — and every throw is logged in the signal box.