By default, Kubernetes moves a pod's state to Running as soon as a pod launches. If the pod is behind a service, the endpoint controller registers an endpoint to Kubernetes immediately. Later on, kube-proxy observes the change of endpoints and configures the host's ipvs or iptables accordingly. Requests from the outside world now go to pods. These operations happen very quickly, so it's quite possible that requests arrive at a pod before the application is ready, especially with bulky software. If a pod fails while running, we should remove it from the pool of a service instantly to make sure no requests reach a bad endpoint.