We migrated our production Kubernetes clusters from Cluster Autoscaler to Karpenter, but with thousands of nodes and hundreds of services, unknown problems were inevitable. This talk covers how we built a rollout strategy where "pause" and "rollback" were as routine as "continue": explicit go/no-go criteria, staged learning with hand-picked canary batches, and tiered rollout with gated alarms. The result: double-digit cost reductions, and a pattern we've reused for many more high-risk migrations.