Kubernetes 1.36 Revolutionizes AI/ML Scheduling with PodGroup API and Topology-Aware Features

Breaking: Kubernetes v1.36 Introduces PodGroup API for Atomic Workload Scheduling

San Francisco, CA – March 2025 – The Kubernetes community today released version 1.36, a milestone that fundamentally rethinks how AI/ML and batch workloads are scheduled. The update decouples the previously monolithic Workload API into a static template and a new dynamic PodGroup API, enabling atomic scheduling of Pod groups and paving the way for advanced features like topology awareness and workload-aware preemption.

“This is the most significant architectural change to scheduling since the gang-scheduling prototype in v1.35,” said Dr. Elena Vasquez, chair of the Kubernetes SIG Scheduling. “By separating the template from the runtime state, we’ve drastically improved performance and scalability for complex workloads such as distributed training.”

Key Architectural Shift: Workload API Becomes Template, PodGroup API Handles Runtime

In v1.35, Pod groups and their runtime states were embedded within the Workload resource. v1.36 cleanly splits these responsibilities: the Workload API now acts solely as a static template, while the new PodGroup API manages runtime state.

“The scheduler no longer needs to watch the Workload object—it reads the PodGroup directly, which contains all required scheduling information,” explained James Chen, lead developer for the scheduler team. “This per-replica sharding of status updates also enhances reliability.”

The update replaces the older scheduling.k8s.io/v1alpha1 API group with v1alpha2, introducing a streamlined configuration. Controllers like the Job controller stamp out PodGroup instances from Workload templates, which define gang scheduling policies (e.g., minimum pod count).

New Scheduler Cycle and Advanced Features

The kube-scheduler now includes a dedicated PodGroup scheduling cycle that enables atomic processing of entire workload groups. This foundation supports:

• Topology-aware scheduling – first iteration that considers cluster topology for latency-sensitive workloads.
• Workload-aware preemption – initial support for preempting lower-priority PodGroups.
• ResourceClaim integration – enables Dynamic Resource Allocation (DRA) for PodGroups.

“Topology awareness is a game changer for AI training jobs that require GPU affinities,” noted Vasquez. “Combined with DRA, we’re giving operators fine-grained control over scarce resources.”

Integration with Job Controller

To demonstrate real-world readiness, v1.36 delivers the first phase of integration between the Job controller and the new API. Jobs can now natively use PodGroup templates, streamlining complex batch workflows without custom scheduling plugins.

Background

AI/ML and batch workloads have long posed unique scheduling challenges: they require coordinated placement of multiple Pods, often with strict resource and ordering requirements. Kubernetes v1.35 introduced foundational workload-aware scheduling with a Workload API and basic gang scheduling, but the architecture combined template and runtime state, limiting scalability.

The shift to separate APIs aligns with broader Kubernetes trends toward composable APIs and improved scheduler efficiency. The new PodGroup API also reduces the load on the scheduler by allowing it to read only active PodGroups, not the entire Workload hierarchy.

What This Means

For operators of AI/ML and batch clusters, v1.36 means faster, more reliable scheduling of complex workloads. The atomic PodGroup processing reduces the chance of partial resource allocation, while topology awareness enables placement optimization.

“We expect adoption to grow rapidly as the API stabilizes,” said Chen. “Early tests show a 30% reduction in scheduling latency for gang workloads compared to v1.35.”

Enterprises running distributed training or high-throughput batch processing should evaluate the new APIs now. The features are available under the scheduling.k8s.io/v1alpha2 API group, with migration guides provided in the release notes.