Kubernetes 1.36 Hits Code Freeze With 80 Enhancements Spanning AI Hardware Scheduling, User Namespaces, and Gateway API Migration
The next Kubernetes release locked its feature set on March 18, delivering 80 enhancements that graduate User Namespaces and four Dynamic Resource Allocation features to GA while introducing gang scheduling and workload-aware placement for AI training clusters.
Kubernetes 1.36 reached code freeze on March 18, locking in 80 enhancements that will ship when the release goes live on April 22. The feature set marks the project’s most significant investment in AI workload scheduling and hardware-aware resource management to date, with 18 features graduating to general availability and 26 new capabilities entering alpha.
The freeze arrives days before KubeCon + CloudNativeCon Europe 2026 opens in Amsterdam on March 23, where the release’s maintainers and contributors are expected to present the technical details.
Dynamic Resource Allocation Reaches Maturity
Four enhancements to the Dynamic Resource Allocation (DRA) API are graduating to GA in this release, cementing Kubernetes’ ability to manage specialized hardware such as GPUs, FPGAs, and custom accelerators. The DRA improvements introduce taints and tolerations for devices, enabling automated pod rescheduling when hardware fails and allowing maintenance windows without cluster disruption.
A new consumable capacity model allows independent resource claims from unrelated pods, even across different namespaces, to allocate specific shares of the same underlying hardware. Device binding conditions and device taints move to beta, giving operators finer control over how workloads interact with physical devices.
These changes directly address a growing operational challenge: as organizations deploy GPU-intensive AI inference and training workloads on Kubernetes, the scheduler must understand not just CPU and memory but the topology and health of accelerator hardware. According to the CNCF, 66 percent of organizations are already running generative AI inference on Kubernetes.
User Namespaces Graduate to Stable
Support for Linux User Namespaces in pods reaches GA status after progressing through alpha and beta over several release cycles. The feature allows processes inside a container to run as root within the pod’s namespace while mapping to unprivileged user IDs on the host, significantly reducing the attack surface for container escape vulnerabilities.
The graduation means the feature is now enabled by default and considered production-ready. For multi-tenant clusters and environments running untrusted workloads, User Namespaces provide a hardening layer that does not require changes to application code.
Gang Scheduling and Workload-Aware Placement
Kubernetes 1.36 introduces new alpha APIs for workload-aware scheduling. A new scheduling.k8s.io/v1alpha2 API defines Workload and PodGroup resources, enabling gang scheduling where all pods in a group must be placed simultaneously or not at all. This is critical for distributed AI training jobs, where a single unscheduled pod in a ring-allreduce topology can leave expensive GPU resources idle.
The release also adds topology-aware workload scheduling, which considers network topology and hardware locality when placing pod groups. New PlacementGenerate and PlacementScore extension points let scheduler plugins evaluate and rank candidate placements for entire workload groups rather than individual pods.
Gateway API Replaces Ingress-NGINX
Kubernetes 1.36 accelerates the migration from Ingress-NGINX to the Gateway API, which the project positions as an expressive, role-oriented successor offering traffic splitting, multi-tenant networking, and Layer 4 and Layer 7 routing. Ingress-NGINX enters best-effort maintenance in this release cycle, meaning it will receive only critical security patches going forward.
OCI Artifacts and Admission Control
OCI Artifacts as a VolumeSource graduates to stable, allowing developers to mount machine learning models and datasets as independent OCI artifacts directly into pods, similar to ConfigMaps. This decouples model data from application containers and reduces the security attack surface.
MutatingAdmissionPolicy also reaches GA, alongside a new manifest-based admission control configuration that lets platform teams secure critical policies using static files on the control plane disk, preventing accidental deletion or circumvention during cluster startup.
Breaking Changes to Watch
Operators upgrading to 1.36 should note that StrictIPCIDRValidation is now enabled by default, meaning Kubernetes will reject IP addresses in legacy octal notation such as “010.000.000.005” and ambiguous CIDR formats. Several metrics have been renamed to conform to naming conventions, including volume_operation_total_errors becoming volume_operation_errors_total. The Service.spec.externalIPs field is now deprecated with warnings.
The beta release (v1.36.0-beta.0) shipped on March 11, with release candidates expected in early April before the final release on April 22.