Canary Release

What is a canary release?

Ship a new model to everyone at once and, if it's broken, everyone suffers together. A canary release is more careful: push the new version to a small slice of users — say 5% — watch the metrics, then grow to 25%, 50%, 100% as confidence builds. The name comes from mining: miners sent a canary down first, and if the bird keeled over, they knew to get out.

How's it different from blue-green deployment?

Blue-green: one big switch
You keep old and new environments side by side, then flip all traffic at once — fast, but the risk lands in that single cut.
Canary: gradual ramp
No single cut. You move traffic to the new version bit by bit, watching as you go, and you can stop anytime.

How do you run one?

Shift traffic by percentage
Use routing or a load balancer to send a set share of users to the new model.
Watch the key metrics
Error rate, latency, conversion — compare new against old, side by side.
Keep a kill switch
The moment metrics sour, pull traffic back to the old version with limited damage.

Why it's the default for shipping AI

You can't fully verify an AI model's "correctness" in a test environment — real users and real data are the final exam. A canary lets the new model face a small, real-world test first, proving itself while protecting almost everyone.

Bottom line: a canary release is "small steps, brakes ready" — let a few people try it, then scale once it's stable.