This comprehensive guide covers Kubernetes GPU operations from driver installation and device plugin internals to resource scheduling, MIG partitioning, time-slicing, DRA migration, monitoring with DCGM, and production troubleshooting with real-world case studies.
Kubernetes 1.36 marks a shift toward native governance of complex AI resources, workload‑aware scheduling, fine‑grained security, and control‑plane scalability, urging platform teams to rethink resource allocation, isolation, and governance as AI workloads move into production.
Kubernetes v1.36 focuses on clearing technical debt rather than adding flashy features, retiring ingress‑nginx, tightening kubelet API auth, optimizing SELinux mounts, externalizing ServiceAccount token signing, expanding DRA for GPU scheduling, graduating MutatingAdmissionPolicy, and removing long‑standing legacy components, all accompanied by a concrete upgrade checklist.
Kubernetes v1.32 introduces 44 enhancements—including 13 stable, 12 beta, and 19 alpha features—spanning dynamic resource allocation, Windows node support, improved kubelet reliability, new API endpoints, and extensive updates to DRA, pod‑level resources, and scheduling, all aimed at strengthening the cloud‑native ecosystem.
Kubernetes v1.32 introduces 44 enhancements—including 13 new stable features, 12 beta features, and 19 alpha features—covering DRA improvements, node and sidecar usability, custom resource selectors, storage updates, scheduling advances, Windows support, and more, with download links and references.
This article explains the design of Kubernetes 1.32 Device Resource Allocation (DRA), detailing the four new CRDs, the roles of kube‑controller‑manager, kube‑scheduler and kubelet plugins, and the required admission webhook and RPC interfaces for managing device resources.
The article examines the scheduling performance and tight coupling issues of Kubernetes DRA before version 1.30, explains the original workflow involving PodSchedulingContext and DRA driver, and then details the latest design that introduces ResourceClaimParameters and ResourceSlice to let the scheduler handle complex device constraints internally.
This article explains how Kubernetes uses the device‑plugin framework to extend resources beyond CPU and memory, details the kubelet registration and allocation workflow, and walks through the relevant source code in pkg/kubelet/cm/devicemanager that builds the OCI spec.
