What is Canary deployment?

A canary deployment is a progressive rollout strategy that limits the blast radius of a release by sending it to a small percentage of traffic before promoting it to the whole fleet. The name comes from canaries in coal mines: the small cohort is the early-warning signal. You deploy the new version alongside the stable one, route a fraction of users (say 1%, then 5%, then 25%) to it, and continuously compare the canary's SLIs, error rate, latency, saturation, against the baseline. If the canary degrades, automation rolls it back before most users ever see the bad version; if it stays healthy, traffic ramps up until the new version owns 100%. Canary deployments are the safer cousin of blue-green deploys and a core technique for spending error budget deliberately rather than betting an entire release on a single big-bang cutover. They depend on good observability: you can only canary what you can measure. For an AI agent driving or supervising a rollout, knowing the canary's current weight, the comparison thresholds, and the history of prior failed canaries, all durable facts worth holding in shared memory, lets it decide whether to promote, hold, or abort.