Canary behavior changes before rollout

Canary plan

Specify the cohort, observation window, success metrics, stop thresholds and rollback or forward-repair procedure. Choose a cohort that exercises the changed path without assigning irreversible risk to unsuspecting users.

Example

A new retrieval ranking runs for a small internal evaluation set first. Compare correct-result selection, empty results, latency and errors with the old ranking on the same queries. Only then consider a limited live cohort if authorized. Do not use production users as an unannounced substitute for missing basic tests.

Decision rule

Set thresholds before seeing the results. If errors cross the stop threshold, halt the expansion and execute the prepared recovery plan. If the sample is too small to judge, report insufficient evidence rather than calling the canary successful because nothing obvious failed.

Acceptance and limits

Inject a known regression in isolation and verify the stop mechanism triggers. Ensure the control and canary cohorts remain identifiable in metrics without logging personal data. This is an original rollout checklist; a clean canary cannot rule out rare failures, long-term effects or unrepresented user paths.