You scaled your gRPC service to five replicas. One of them is on fire. The other four are asleep. This isn’t a bug — it’s a direct consequence of how Kubernetes moves traffic between pods, and once you see the mechanism, the fix is obvious.