A canary deployment limits the initial production exposure of a change, then uses real-traffic evidence before wider promotion. It turns deployment into a validation step rather than treating a successful change as proven. section 9.1.3 section 10.6.2
Use a canary as part of a disciplined release loop:
- Deploy the changed component to a limited production scope. section 9.1.3 section 10.6.2
- Observe production metrics and compare them with the expected behavior. section 9.1.3 section 10.6.2
- Promote only when the observed evidence supports the change; otherwise investigate or stop the rollout. section 9.1.3 section 10.6.2
Model-tier change
After swapping an AI model tier, a team reviews the change, validates it with real traffic through a canary, and checks for quiet regressions before treating the optimization as shipped. section 9.1.3










