# Review what happens when the release metrics disappear

Analysis failures are part of the rollout design. An empty or broken query must have an explicit outcome before it controls production exposure.

By Cobnex editorial. Published 2026-09-10. Updated 2026-09-11.

## Inspect the query's meaning

Check the release labels, operation filters and observation window. Confirm that the query selects candidate traffic and the behaviour being assessed.

A query returning a plausible number can still inspect the stable release or an unrelated endpoint. Compare it with known synthetic requests before relying on it.

Keep the denominator visible for rate calculations and define how zero observations are handled.

## Test every analysis outcome

Supply healthy, unhealthy, empty and error results through a safe test. Confirm the controller's action for each.

Missing data should follow the documented hold or review policy. It should not become success because an expression treats an empty collection as passing.

Also test late data and repeated provider failures. The release should not remain indefinitely ambiguous without an owner or timeout policy.

## Check the action beyond status

Verify what abort, pause and promotion do to traffic, instances and background work. A named state in the controller is useful only when its operational effect is understood.

Inspect in-flight requests and candidate-produced data after abort. Recovery may need application reconciliation in addition to routing changes.

Confirm the stable path's capacity and compatibility before describing abort as a safe response.

## Review the human override

Define who can override an inconclusive analysis and what evidence they must inspect. Keep the action recorded with the release decision.

An override should not erase the failed measurement or make the report appear automatically healthy. Preserve the reason and any remaining limitation.

Approve the rollout policy when its observations and actions are testable end to end. The important review question is whether the system reaches the right decision with imperfect evidence, since that is the condition it will eventually encounter in production.

## Exercise a rare path before expanding traffic

Ask for a candidate that succeeds on ordinary requests but fails a low-volume critical path, such as an approval with an attachment. An aggregate error rate may barely change. The rollout needs a targeted synthetic journey or segmented signal capable of detecting that failure before expansion.

Then inspect the rollback target after the candidate has written data. If the old version cannot read the new representation, shifting traffic back is not a complete rollback. The review should require a compatible write strategy or another explicit recovery plan.

## Sources

- [Argo Rollouts: analysis](https://argo-rollouts.readthedocs.io/en/stable/features/analysis/)