Every team eventually ships a release that breaks production, and the first job is to restore service, not to debug live. A rollback returns to the last known good version, by redeploying the previous artifact, undoing a Kubernetes rollout or switching blue-green back. Plan it before you need it, with immutable versions and database changes that old code can still handle. Rolling forward with a quick fix is fine for small bugs, but slower when things are badly broken.