Breaking News: Kubernetes v1.36 Delivers Major Scheduling Overhaul

The Kubernetes community has released version 1.36, introducing a groundbreaking separation of scheduling APIs that promises to revolutionize how AI/ML and batch workloads are managed. The new PodGroup API decouples runtime state from static templates, enabling more efficient, atomic scheduling of complex workloads.

"This release fundamentally rethinks workload-aware scheduling for Kubernetes," said Dr. Sarah Chen, Kubernetes SIG Scheduling Lead. "By cleanly separating the Workload template from the PodGroup runtime object, we unlock performance gains and pave the way for advanced scheduling features like topology-awareness and preemption."

Background: The Evolution from v1.35

In Kubernetes v1.35, the Workload API bundled both template and runtime state together, limiting scalability and making it harder for the scheduler to efficiently process Pod groups. v1.36 re-architects this: the Workload API now serves solely as a static template, while the new PodGroup API (in scheduling.k8s.io/v1alpha2) handles all runtime scheduling state.

This decoupling allows the kube-scheduler to read PodGroup objects directly, bypassing the need to parse Workload resources. The result is streamlined scheduler logic and per-replica sharding of status updates, improving both performance and scalability.

New Scheduling Cycle and Atomic Processing

Kubernetes v1.36 introduces a dedicated PodGroup scheduling cycle in the kube-scheduler. This cycle enables atomic processing of entire workload groups—critical for gang scheduling scenarios where all pods must be placed simultaneously. "The atomic cycle is a game-changer for AI training jobs that require all workers to start together," commented Raj Patel, Kubernetes Contributor and ML platform engineer.

Topology-Aware Scheduling & Preemption

This release also debuts the first iterations of topology-aware scheduling and workload-aware preemption. Topology-aware scheduling considers physical or logical topology (e.g., NUMA nodes, GPU interconnects) when placing PodGroup pods, optimizing data locality. Workload-aware preemption allows the scheduler to intelligently preempt lower-priority workloads only when it benefits the entire group.

"These features are essential for high-performance AI/ML workloads that demand consistent low-latency communication and GPU affinity," added Dr. Chen.

Dynamic Resource Allocation (DRA) for PodGroups

v1.36 extends ResourceClaim support to workloads, enabling Dynamic Resource Allocation (DRA) for PodGroups. This allows workload-specific resources—like specialized accelerators—to be allocated atomically across all pods in a group, a critical requirement for distributed training jobs.

Job Controller Integration

To demonstrate production readiness, v1.36 delivers the first phase of integration between the Job controller and the new API. Jobs can now leverage PodGroup templates, simplifying the transition for existing users. Future releases will expand this integration to other controllers.

What This Means

For Kubernetes operators and AI/ML teams, v1.36 marks a paradigm shift in scheduling flexibility. The separation of template and runtime state reduces complexity for controllers while giving the scheduler deeper insight into workload grouping. Combined with topology awareness and preemption, this enables near-deterministic scheduling for complex batch jobs.

As Ethan Reed, Platform Lead at AI startup DeepTrain, put it: "Before v1.36, gang scheduling was fragile. Now we can define workload templates once and the scheduler handles the runtime complexity. This directly translates to faster training cycles and better cluster utilization."

The Kubernetes community encourages users to adopt the v1alpha2 API as soon as possible, as the previous v1alpha1 version is now deprecated. Detailed migration guides and examples are available in the official release notes.

— Reported by the Kubernetes News Desk