Kubernetes Blog
Follow
Kubernetes v1.36: Pod-Level Resource Managers (Alpha)
Kubernetes v1.36 introduces Pod-Level Resource Managers as an alpha feature, enhancing resource management for performance-sensitive workloads. It extends kubelet's Topology, CPU, and Memory Managers to a pod-centric resource allocation model, moving beyond per-container specifications. This addresses the challenge of providing exclusive, NUMA-aligned resources for primary application containers while supporting lightweight sidecars efficiently. Previously, achieving predictable performance often meant allocating exclusive resources to all containers, which was wasteful for sidecars. Alternatively, not doing so sacrificed the pod's Guaranteed QoS. Pod-level resource managers enable hybrid allocation, allowing high-performance workloads to achieve NUMA alignment without wasting resources. For example, a latency-sensitive database pod can have its main container receive exclusive CPU and memory, while sidecars share a distinct pod shared pool, isolated from other node resources. Another use case involves ML workloads where the training container gets exclusive NUMA-aligned resources, and a service mesh sidecar runs in the general node-wide shared pool. CPU isolation is managed by disabling CFS quota enforcement for exclusive containers and enforcing it at the pod level for shared pool containers. Enabling requires specific kubelet feature gates, Topology Manager policies, and static CPU and Memory Manager configurations. New kubelet metrics provide observability into resource allocations and container assignments. This feature is currently in alpha, with known limitations and caveats, and user feedback is encouraged through Kubernetes community channels.
