What is Canary Release?
Also called canary deployment, progressive rollout.
A canary release is a deployment strategy that sends a new version to a small share of traffic first, compares its behavior against the existing version, and expands the rollout only if the metrics look healthy. It limits the number of users exposed to a defective release and provides real production signal before full exposure. Its usefulness depends entirely on having metrics good enough to detect the problem.
A typical rollout moves through increasing shares, for example one percent, then ten, then fifty, then all traffic, pausing at each step long enough to gather a meaningful sample. Automated analysis compares error rates, latency, and business metrics between the two versions and either promotes to the next step or rolls back. Manual canaries follow the same shape with a human reading dashboards.
The technique differs from a blue green deployment, where two full environments exist and traffic switches over all at once. Blue green gives an instant switch and an instant way back but exposes everyone simultaneously. Canary limits exposure but requires both versions to run side by side against the same data, which constrains schema and contract changes to backward compatible ones.
Sample size is the practical limit. A one percent canary on modest traffic may take hours to accumulate enough requests to distinguish a real regression from noise, and a rare failure affecting one flow in a hundred may never appear in the canary window at all. Canaries reliably catch broad regressions, and are much weaker on narrow or slow burning ones.
Selecting canary traffic requires thought. Random selection risks exposing important customers, while routing only internal users may miss the configurations that break. Sticky assignment matters too: a user bouncing between versions mid session can hit inconsistent behavior. Many teams route by account so each account sees one version consistently for the duration of the rollout.
Key points
- Exposes a small traffic share first, then expands on healthy metrics.
- Both versions run at once, so changes must be compatible.
- Blue green switches all traffic; canary ramps gradually.
- Low traffic canaries lack the samples to catch rare regressions.
- Route by account so a user does not flip between versions.
In practice
A new agent execution path rolls out to five percent of workspaces. Over two hours the automated comparison shows equal error rates but a fifteen percent rise in ninety fifth percentile task duration in the canary group. The rollout pauses rather than promoting, an added retry loop is identified as the cause, and the corrected build ramps through five, twenty five, and one hundred percent the following day without regression.