The blog post discusses a new feature in Kubernetes that allows for in-place resource adjustments without the need to recreate pods, benefiting AI training workloads and edge computing deployments. This feature directly modifies cgroup settings on running containers, eliminating the need for maintenance windows and preventing applications from going offline during transitions.
One of the key highlights of the new release is the introduction of gang scheduling, a preview feature designed to support distributed AI workloads that require multiple pods to start simultaneously. This capability, also known as KEP-4671, allows for coordinated scheduling of pods to ensure seamless operation for distributed applications.
To utilize these new features, users must have cgroups v2 enabled on their Linux nodes, as Kubernetes 1.35 has deprecated support for cgroups v1. While most enterprise Linux distributions already support cgroups v2, older deployments may require OS upgrades to take advantage of the enhanced resource management capabilities offered by Kubernetes.
