When a worker node stops responding, the Service keeps routing traffic only to pods that still report Ready. The pods on the missing node drop out of the endpoint list, so the application runs at reduced capacity but stays up.
The deployment controller notices it is short of replicas and schedules replacements on the nodes that still have spare capacity. Once those new pods pass their readiness checks they join the Service, and full capacity returns. The failed node is left isolated for investigation rather than repaired in place.
This is a scripted illustration, not a live cluster. It assumes replicated workloads and spare capacity on the remaining nodes, and it compresses timing that would take longer in practice. Paging does not send a real notification.
