Rolling a change out to a small slice of real users first, to catch problems before they affect everyone.
Example: Shipping a change to 1% of users before the full rollout.
In practice: The canary's health — error rates, latency — is monitored automatically, and the rollout halts or reverses if it looks worse than the baseline population.