Retire models on purpose

How to do it

Old models left running keep costing money and keep a door open. Deleted models break every appeal about a decision they made. Mark the version as retired in the registry with a date and a reason. Find the last decision that names it, as in keep a record of every decision about a person. Delete its weights and snapshot only after that record's retention ends.

Examples

Python example

def retire(version: ModelVersion, reason: str) -> None:
    if serving.endpoints_using(version):
        raise StillServing(version.id)
    registry.set_tag(version, "retired", f"{date.today().isoformat()}: {reason}")
    feature_jobs.stop_for(version)

TypeScript example

export async function retire(version: ModelVersion, reason: string) {
  if ((await serving.endpointsUsing(version)).length) throw new StillServing(version.id);
  await registry.setTag(version, 'retired', `${new Date().toISOString().slice(0, 10)}: ${reason}`);
  await featureJobs.stopFor(version);
}

How to check

No serving endpoint or alias points at a version marked retired. Every retired version has a date and a reason. A query finds no decision record that names a version whose weights are gone.

Sources

• MLflow Model Registry

Related rules

• MLO-21 Keep every model in a registry with versions and aliases in Machine learning
• MLO-38 Keep a record of every decision about a person in Machine learning