How would you perform a canary release and decide whether to continue or roll back?

Deploy the new version to a small, representative portion of traffic and compare it with a stable baseline. Define decision thresholds for error rate, latency, resource use, and relevant business outcomes before starting. Check that telemetry distinguishes versions and that database changes remain compatible. Increase exposure in stages with enough observation time for the workload. Roll back or stop on clear regressions, and account for low traffic or delayed failures. A canary is informative only if routing, measurement, and the rollback path have been tested.