Releasing carefully
A canary rollout sends a small slice of live traffic, perhaps one percent, to a new model version while the rest stays on the proven one. If the canary stays healthy, traffic ramps up gradually. If it misbehaves, the blast radius was tiny.
What to watch on the canary
- Quality and proxy metrics compared against the stable model.
- Operational health like latency, errors, and resource use.
- Prediction distribution to catch surprising output changes early.
Ramp strategy
- Increase share in stages, such as one then five then twenty five percent, pausing to observe at each step.
- Define clear promotion and rollback criteria before starting.
- Keep the comparison fair by routing similar traffic to both.
Why it beats a big bang
A full switch exposes every user to an untested model at once. Canary keeps failures cheap and reversible, the core safety property of progressive delivery.
Key idea
A canary rollout routes a small traffic slice to a new model and ramps only if metrics stay healthy, keeping any bad release cheap and reversible.