Kubernetes 1.35 features that change Day 2 operations

"In-place pod resource update is the standout GA feature of the latest Kubernetes release. In prior versions, modifying CPU, memory requests, and limits on a container forced a pod restart, disrupting stateful workloads and long-running processes. With the in-place pod vertical scaling feature, you can resize the resources of a running pod's containers seamlessly without killing the pod. Implementing this capability is similar to manipulating the specifications of an existing pod."

"By using kubectl patch or kubectl edit, you can directly change the CPU and memory requests. When manipulating controllers such as deployments and StatefulSets, the spec can be changed and applied to make it persistent. Remember that this capability is limited to CPU and memory, not other resources such as ephemeral storage, which still forces a pod restart. This feature has a significant impact on pod vertical autoscaling."

"When integrated with Vertical Pod Autoscaling (VPA), a pod can be vertically scaled based on real-time metrics from sources such as the Metrics Server or Prometheus. It's especially useful for managing stateful services, where restarts trigger data rebalancing or failovers. For AI and machine learning (ML) workloads, it helps preserve in-memory caches and model checkpoints, minimizing interruptions during training, fine-tuning, and inference."

Kubernetes 1.35 introduces generally available enhancements that improve workload reliability and efficiency. In-place pod resource update allows changing CPU and memory requests and limits on running containers without restarting pods. Resource updates can be applied via kubectl patch or kubectl edit and persisted when updating controller specs such as Deployments and StatefulSets. The capability is limited to CPU and memory; changes to other resources like ephemeral storage still require pod restarts. Integration with Vertical Pod Autoscaling enables vertical scaling based on real-time metrics from sources such as Metrics Server or Prometheus. The feature benefits stateful services and AI/ML workloads by preserving in-memory caches and model checkpoints.

Read at thenewstack.io