Canary deployment gradually rolls out a new model to small percentage of traffic, monitoring for problems before full release. Process: Deploy new model to 1-5% of traffic, monitor key metrics, if healthy increase to 10%, 25%, 50%, 100%. Rollback if issues. Why canary: Limits blast radius of problems. Real traffic validation. Quick rollback path. Builds confidence incrementally. Metrics to monitor: Latency, error rates, business metrics, user feedback. Compare canary to control population. Automation: Progressive delivery tools (Flagger, Argo Rollouts) automate percentage increases based on metrics. Rollback triggers: Error rate spike, latency increase, business metric degradation, manual halt. Automatic rollback possible. Traffic routing: Load balancer routes percentage to new model deployment. Consistent routing (same user always sees same version) often preferred. Duration per stage: Hours to days depending on traffic volume and confidence requirements. Comparison to A/B testing: Canary is for safe rollout, A/B is for choosing between versions. Can combine: canary new model, then A/B test features. Best practice: Start small, automate monitoring, have instant rollback ready.