A canary deployment releases a change to a small subset of traffic first, monitors for errors, and then gradually expands if the release behaves as expected.
A canary deployment shifts a small percentage of user requests to the new version while the majority remain on the stable version. Automated checks verify metrics such as error rate, latency, and business outcomes before increasing traffic.
That gradual exposure limits blast radius when a release contains a defect. Instead of affecting every user, only the canary group sees the problem. The release can be halted or rolled back automatically before it spreads. The downside is routing complexity and the need for accurate health signals that distinguish normal variance from real failure.
Think of it like this. Think of miners carrying a canary into a coal mine. Only a small sample enters first. If the canary survives, the mine is considered safe for everyone else.
The deployment system exposes both the old and new versions simultaneously. Traffic is split by weight, header, or user segment. Observability tools compare error rates, latency, and resource usage between canary and baseline. If metrics stay healthy, traffic weight increases in steps until the rollout is complete.
"Canary means 5 percent always works." The right percentage depends on traffic volume and detection sensitivity. "Automated promotion means no monitoring." Automated canaries still need meaningful thresholds and alerting. "Canary replaces rollback plans." Fast rollback remains necessary when automated checks miss subtle failures.
Reduces release risk and improves confidence, but adds routing complexity and requires reliable metrics. Works best with observability-backed pipelines; less valuable for low-traffic systems where failure detection is slow.