Recover an interrupted model alias operation

When you need it

Looking for a normal rollback after a completed promotion? Follow the training_report and Run now steps.

An AliasOutcomeUnknownError or a pending alias event means a registry write may have succeeded before its response was received. Further controlled training, promotion and champion scoring can remain blocked until that event is reconciled. This procedure concerns registry aliases; a failed score or dashboard refresh after a confirmed promotion does not by itself require alias recovery.

flowchart LR
    A["Stop every alias writer"] --> B["reconcile_alias_change"]
    B --> C["Read all four roles and saved intent"]
    C --> D["committed: finish metadata"]
    C --> E["aborted: nothing changed"]
    C --> F["AliasConflictError: inspect"]

What to edit

First stop every job and manual operation that can write the model's aliases. Confirm the original run has terminated and cannot retry. Preserve its inputs, logs, model/version tags and all four role assignments: champion, challenger, previous_champion, previous_challenger.

Obtain the original destination version and event ID from the failed operation's logs, pending_alias_event and promotion_<event_id>. New operations save their intent in alias_intent plus numbered chunks before publishing the pending marker. Run recovery with the same Skyulf Core version as the jobs that wrote the intent. Use the same admission authority as the original writers; do not invent a new lock or clear the pending marker just to make a job run.

The command

The MLflow registry recovery procedure contains the SDK call with its required admission, destination version and event ID. Run that call from an authorized operator environment while writers remain stopped. There is no Bundle lifecycle_action=recover convenience action. The recovery API validates observed state against the saved intent; it does not set or delete aliases for you.

For a normal candidate decision whose outcome is known, use the train job's approve/reject/rollback actions. Do not substitute rollback for reconciliation of an unknown outcome.

What you should see

Result	Meaning	Next step
committed	Intended role changes already exist; metadata is completed and verified, then this pending marker is cleared.	Read the controlled champion and resume the appropriate downstream operation.
aborted	Roles and owned metadata still match their starting values; no promotion is completed.	Review and retry the original business operation with fresh checks if still wanted.
AliasConflictError	Mixed roles, unexpected changes or contradictory metadata prevent automatic reconciliation.	Keep writers stopped and inspect the complete intent and observed state.

An unchanged retry of an already completed repair returns the same outcome. Unchanged aliases alone do not prove completion of metadata-only actions such as rejection. Recovery does not undo prediction rows, endpoint traffic or tables.

Common errors and fixes

Symptom	Fix
Another writer still owns admission	Verify that exact owner's run state; follow the registry guide's conditional recovery procedure. Never release a different owner's claim.
Only some roles changed	Compare every role and marker with the saved plan under exclusive operator control; follow the documented manual procedure before retrying reconciliation.
Alias intent digest differs from its stored chunks and the alias_intent manifest shows "encoding":"json-framed-v1"	The operator environment runs an older Core that cannot read framed chunks. This is a version mismatch, not corruption. Install the writers' Core version and retry.
An old event has no complete intent	Automatic recovery is unavailable. Restore from verified original evidence with the platform owner; keep the model blocked if that evidence is insufficient.
An old intent was replaced by a newer operation	The latest snapshot cannot repair the earlier event. Use the original durable evidence.
A tag budget error occurs	Preserve current control receipts and pending evidence; review the registry guide before retiring obsolete audit tags.

Return to model lifecycle decisions only after reading back a consistent controlled state.