Kubernetes v1.35 introduces the RestartAllContainers action (graduating to beta in v1.36), enabling in-place pod restarts without recreating the entire Pod object. By keeping the Pod sandbox intact — preserving network identity, GPU/TPU bindings, and storage mounts — containers can recover from failures while avoiding control plane churn, scheduling delays, and resource race conditions. This is especially valuable for large batch and AI/ML workloads where pod recreation at scale causes 'Thundering Herd' scheduling storms. The feature is configured via restartPolicyRules in the pod spec, introduces an AllContainersRestarting condition for observability, and requires applications to be idempotent and handle abrupt SIGKILL termination. Recovery times for JobSet workloads reportedly dropped from minutes to seconds.
