Loading
Rolling out a change to a small percentage of users first to catch problems before everyone sees it.
Route 1-5% of traffic to the new version; monitor error rates and latency. If metrics are healthy, gradually increase to 100%. If not, route all traffic back to the old version. Lower risk than big-bang deploys.
Requires traffic splitting (weighted routing or header-based), comparable monitoring between canary and baseline, and automated rollback triggers. Feature flags are a code-level equivalent. Canary + progressive delivery = the state of the art for safe continuous deployment.