When this happens, the service is almost certainly selecting the wrong pods.

Kubernetes services don’t care about deployments or rollout status. They route traffic purely based on label selectors. If your new pods have labels that don’t exactly match what the service expects, traffic will continue flowing to the old ReplicaSet even though the rollout completed successfully.

This often happens after small refactors where labels are renamed or reorganized, and the service definition isn’t updated accordingly.

Takeaway: If traffic isn’t shifting, always check service selectors before blaming the rollout