Kubernetes v1.37: Advancing Workload-Aware Scheduling

Kubernetes v1.37 promoted its core Workload and PodGroup APIs, gang scheduling, Workload-Aware Preemption, and shared DRA ResourceClaims for PodGroups to Beta, with the Workload and PodGroup APIs moving to v1beta1, one step from General Availability. The release introduces the new CompositePodGroup API for multi-level topology constraints, gang scheduling, and preemption policies, plus controller integration APIs and the workloadbuilder Go library, and replaces v1alpha2 with v1alpha3 in a breaking change around disruptionMode. The minCount field is now mutable for elastic workloads, PodGroups are queued as a single top-level object rather than per-Pod, and the WorkloadAwarePreemption feature gate merged into GenericWorkload.

AI/ML and complex batch workloads continue to push the boundaries of Kubernetes scheduling. Following the foundational workload-centric enhancements introduced in previous releases, Kubernetes v1.37 delivers the next major milestone in the Workload-Aware Scheduling WAS journey. In this release, the core Workload and PodGroup APIs—enabling gang scheduling—along with Workload-Aware Preemption WAP and shared DRA ResourceClaims for PodGroups, all graduate to Beta, solidifying their role in the Kubernetes ecosystem. To address the hierarchical scheduling requirements of modern high-performance distributed workloads, v1.37 introduces the new CompositePodGroup https://kubernetes.io/docs/concepts/workloads/compositepodgroup-api/ API. This new API allows expressing multi-level topology constraints, gang scheduling, and preemption policies for complex, heterogeneous groups of Pods. Crucially, this architectural expansion unlocks native scheduling support for advanced workload structures commonly managed by higher-order extension APIs such as JobSet and LeaderWorkerSet LWS . Alongside these API additions, v1.37 focuses on streamlining adoption by introducing a new set of controller integration APIs and the workloadbuilder Go library. These provide standardized building blocks that significantly simplify how out-of-tree controllers can integrate with WAS capabilities. Utilizing these new tools, the native Job controller integration has been upgraded to fully consume the expanded WAS APIs—enabling advanced scheduling policies, flexible disruption modes, and topology-aware scheduling for standard batch workloads. Kubernetes v1.37 delivers a major milestone: Workload https://kubernetes.io/docs/concepts/workloads/workload-api/ / PodGroup https://kubernetes.io/docs/concepts/workloads/podgroup-api/ APIs and gang scheduling are officially graduating to Beta. This graduation signals that native, "all-or-nothing" scheduling for workloads is solidifying for wider adoption. Key updates to the API and gang scheduling algorithm in this release include: The core Workload and PodGroup APIs have been promoted to v1beta1, meaning they are now one step away from General Availability GA . For early adopters who have been testing these features, take note of the alpha versioning transition: v1alpha2 has been entirely replaced by v1alpha3. This transition introduces breaking changes designed to clean up the API structure around disruptionMode . A significant under-the-hood improvement in v1.37 makes the PodGroup a first-class citizen in the scheduling queue. Previously, even if belonging to a PodGroup, all member Pods were queued individually. Now, only the top-level PodGroup object is queued. This ensures all Pods share the same queueing behavior and lays the groundwork for more advanced PodGroup queueing strategies in the future. In earlier iterations, the minCount field, which dictates the minimum number of Pods required to successfully schedule a PodGroup, was strictly immutable. In v1.37, minCount is now mutable. This API change unlocks flexibility for elastic workloads. Controllers can now dynamically adjust the minimum required size of a gang on the fly, allowing workloads to gracefully degrade or expand without interrupting already-scheduled Pods. In Kubernetes v1.37 the separate WorkloadAwarePreemption feature gate https://kubernetes.io/docs/reference/command-line-tools-reference/feature-gates/ for workload-aware preemption was merged into the GenericWorkload feature gate, becoming a core part of the gang scheduling effort. While the core concepts of workload-aware preemption stay the same, there are some differences between the v1.36 and v1.37 releases: To check whether a preemptor can fit in the cluster thanks to preemption, the scheduler simulates the removal of all potential victims and re-runs the scheduling algorithm. After that it tries to reprieve as many victims as possible. In the v1.36 release, the scheduling algorithm was run for each victim reprieval, verifying whether with the victim reprieved, the algorithm can still find a valid placement for the preemptor. In v1.37, the scheduling algorithm is run only once and the preemptor Pods are assumed based on its output. Later, the reprieval checks whether a victim can still run in its place with the preemptor assumed. One of the limitations of v1.36 was the fact that the default preemption for single Pods was not aware of PodGroups and was not respecting their disruptionMode fields, allowing for disruption of single Pods even when the PodGroup had disruptionMode: {all: {}} set. Kubernetes v1.37 removes this limitation; the default preemption now respects the PodGroup disruptionMode field. disruptionMode fields During the promotion of the API to Beta, the disruptionMode field was changed to decouple its naming from the PodGroup object, allowing consistent naming across PodGroups and CompositePodGroups. The modes changed as follows: PodGroup became all , and Pod became single . preemptionPolicy In v1.36, the PodGroup does not have a preemptionPolicy field. The PodGroup can perform preemption as long as none of the Pods forming it has preemptionPolicy: Never set. In v1.37, when the PodGroupPreemptionPolicy https://kubernetes.io/docs/reference/command-line-tools-reference/feature-gates/ PodGroupPreemptionPolicy feature gate is enabled, a PodGroup also has a preemptionPolicy field. It serves as an authoritative field for whether a PodGroup can perform preemption. In Kubernetes v1.36, workload-aware scheduling established a clean separation between static workload templates Workload and runtime group state PodGroup , but the supported scheduling policies were limited to a single, flat group. The CompositePodGroup API, introduced in Kubernetes v1.37, extends this model to support hierarchical scheduling requirements. This API allows its consumers to express multi-level scheduling requirements by organizing a workload in a tree-shaped hierarchy consisting of CompositePodGroup and PodGroup objects. Each CompositePodGroup carries policies and constraints that apply to other groups CompositePodGroups and/or PodGroups , similar to how PodGroups govern scheduling behavior for a flat group of Pods. The scheduler treats such a hierarchy as a single scheduling unit and aims to satisfy the requirements specified by every group within that hierarchy. To express multi-level scheduling requirements, you define a hierarchy of templates in a Workload object. Controllers then create the corresponding CompositePodGroup and PodGroup objects from that hierarchy. To support this, the Workload API is extended with the spec.compositePodGroupTemplates field. Each CompositePodGroupTemplate defines a template for a parent CompositePodGroup and directly nests the templates podGroupTemplates and/or compositePodGroupTemplates from which its child groups derive. Below is a sample Workload object that defines a two-level template hierarchy: apiVersion: scheduling.k8s.io/v1beta1 kind: Workload metadata: name: example-workload annotations: kubernetes.io/description: "Two-level workload hierarchy requiring 4 worker Pods and 1 driver Pod to schedule together." spec: compositePodGroupTemplates: - name: workload-root schedulingPolicy: gang: minGroupCount: 2 podGroupTemplates: - name: workers schedulingPolicy: gang: minCount: 4 - name: driver schedulingPolicy: gang: minCount: 1 After creating example-workload , a controller can stamp out the corresponding runtime group objects from these templates: A root CompositePodGroup that references the workload-root template in example-workload and carries its group-level scheduling policy gang scheduling with minGroupCount: 2 : apiVersion: scheduling.k8s.io/v1alpha3 kind: CompositePodGroup metadata: name: example-root-group annotations: kubernetes.io/description: "Root group coordinating gang scheduling across child worker and driver PodGroups." spec: workloadRef: workloadName: example-workload templateName: workload-root schedulingPolicy: gang: minGroupCount: 2 Two child PodGroup objects example-workload-workers and example-workload-driver that reference their respective leaf templates in example-workload and link to the root group via parentCompositePodGroupName : apiVersion: scheduling.k8s.io/v1beta1 kind: PodGroup metadata: name: example-workload-workers annotations: kubernetes.io/description: "Worker group requiring at least 4 Pods to be scheduled together." spec: parentCompositePodGroupName: example-root-group workloadRef: workloadName: example-workload templateName: workers schedulingPolicy: gang: minCount: 4 --- apiVersion: scheduling.k8s.io/v1beta1 kind: PodGroup metadata: name: example-workload-driver annotations: kubernetes.io/description: "Driver group requiring 1 Pod to schedule alongside the workers." spec: parentCompositePodGroupName: example-root-group workloadRef: workloadName: example-workload templateName: driver schedulingPolicy: gang: minCount: 1 To schedule a hierarchical workload, kube-scheduler evaluates the entire group tree as a unified scheduling unit: minGroupCount of child groups when using the gang policy , while each leaf PodGroup must satisfy its own Pod-level policy for example, placing at least minCount of member Pods when using the gang policy . Kubernetes v1.37 extends workload-aware preemption to support CompositePodGroup hierarchies as well. Specifically, if a CompositePodGroup cannot be scheduled due to insufficient capacity in the cluster, the scheduler can invoke preemption to evict lower-priority workloads in order to fit the Pods belonging to that CompositePodGroup. A CompositePodGroup can be selected for preemption as well. To specify the desired behavior during preemption, workload owners can specify an appropriate disruptionMode in the CompositePodGroup spec: single all In Kubernetes v1.37, topology-aware scheduling expands to support complex, multi-level workload hierarchies and delivers performance improvements for existing single-level deployments. In Kubernetes v1.36, we introduced foundational topology-aware scheduling, allowing you to define co-location constraints directly on a PodGroup. While effective for single-level groupings, complex distributed workloads—such as large-scale AI/ML training, JobSet deployments, or disaggregated inference via LeaderWorkerSet LWS —often require co-location across multiple levels of cluster infrastructure simultaneously. For example, an entire workload may need to run within a single availability zone, while different parts of that workload such as specific worker groups or driver processes require strict co-location within specific server racks. In Kubernetes v1.37, alongside the new CompositePodGroup API scheduling.k8s.io/v1alpha3 , topology-aware scheduling expands to support multi-level topology-aware scheduling . You can now express complex co-location requirements by specifying topology constraints at different levels of a group hierarchy. During hierarchical scheduling, the kube-scheduler resolves multi-level topology constraints in a top-down manner. Specifically, topology domains that are considered during the scheduling of a child group are confined within a topology domain that corresponds to the placement assumed by the parent group. Using the updated Workload API scheduling.k8s.io/v1beta1 , you can configure multi-level topology constraints directly within compositePodGroupTemplates . In the example below, the parent template constrains the overall workload to a single availability zone topology.kubernetes.io/zone , while child templates for workers and driver constrain their respective Pods to server racks topology.example.com/rack within that selected zone: apiVersion: scheduling.k8s.io/v1beta1 kind: Workload metadata: name: multi-level-tas-workload namespace: job-ns annotations: kubernetes.io/description: "Workload defining zone-level co-location for the root group and rack-level co-location for child groups." spec: compositePodGroupTemplates: - name: root schedulingPolicy: gang: minGroupCount: 2 schedulingConstraints: topology: - key: topology.kubernetes.io/zone podGroupTemplates: - name: workers schedulingPolicy: gang: minCount: 8 schedulingConstraints: topology: - key: topology.example.com/rack - name: driver schedulingPolicy: gang: minCount: 1 schedulingConstraints: topology: - key: topology.example.com/rack When a controller creates an instance of this workload at runtime, it spawns the corresponding runtime objects from these templates: root template, carrying the availability zone topology constraint and the hierarchical gang scheduling policy. tas-workload-workers and tas-workload-driver , each referencing the root CompositePodGroup as their parent group via the parentCompositePodGroupName spec field: apiVersion: scheduling.k8s.io/v1alpha3 kind: CompositePodGroup metadata: name: tas-workload-root namespace: job-ns annotations: kubernetes.io/description: "Root group constraining the entire workload to a single availability zone." spec: workloadRef: workloadName: multi-level-tas-workload templateName: root schedulingPolicy: gang: minGroupCount: 2 schedulingConstraints: topology: - key: topology.kubernetes.io/zone --- apiVersion: scheduling.k8s.io/v1beta1 kind: PodGroup metadata: name: tas-workload-workers namespace: job-ns annotations: kubernetes.io/description: "Worker group requiring 8 Pods co-located within a single rack in the selected zone." spec: parentCompositePodGroupName: tas-workload-root workloadRef: workloadName: multi-level-tas-workload templateName: workers schedulingPolicy: gang: minCount: 8 schedulingConstraints: topology: - key: topology.example.com/rack --- apiVersion: scheduling.k8s.io/v1beta1 kind: PodGroup metadata: name: tas-workload-driver namespace: job-ns annotations: kubernetes.io/description: "Driver group requiring 1 Pod placed in a rack within the selected zone." spec: parentCompositePodGroupName: tas-workload-root workloadRef: workloadName: multi-level-tas-workload templateName: driver schedulingPolicy: gang: minCount: 1 schedulingConstraints: topology: - key: topology.example.com/rack During scheduling, the scheduler evaluates multiple candidate availability zones across the cluster for tas-workload-root . For each candidate zone, it subdivides the nodes by rack topology to explore feasible rack placements for tas-workload-workers and tas-workload-driver strictly within that zone, systematically evaluating multiple combinations across available zones and racks before making a scheduling decision. By allowing topology constraints to be modeled hierarchically, Kubernetes v1.37 provides a structured way to express multi-level co-location requirements across complex cluster infrastructures. Alongside the Alpha introduction of multi-level hierarchies, Kubernetes v1.37 reduces the cost of placement evaluation for existing single-level topology-aware scheduling. We are continuously working to optimize the efficiency of placement evaluation algorithms in kube-scheduler and plan to deliver further performance improvements in future releases. Kubernetes v1.37 introduces new standard building blocks so that every controller can expose the same scheduling primitives in their own APIs, and share the same logic for translating them into scheduling objects. These primitives express specific scheduling behaviors — such as policies or disruption logic — while leaving the field naming flexible for each controller. A prime example of this is the native Job controller, which we detail in the next section. Types prefixed with WorkloadPodGroup describe a leaf group of Pods; types prefixed with WorkloadCompositePodGroup describe a group of groups. A controller embeds them verbatim into its own API, under whatever field name fits its domain: WorkloadPodGroupSchedulingPolicy — either basic , meaning standard Pod-by-Pod scheduling, or gang with a WorkloadPodGroupSchedulingConstraints — the topology constraints topology .key the group's Pods must be co-located within. WorkloadPodGroupDisruptionMode — WorkloadPodGroupResourceClaim — the ResourceClaims shared across the group. Only the shapes are shared, so controllers retain full autonomy over how they name and nest these fields in their own APIs. The workloadbuilder library turns that intent into the scheduling objects. A controller describes its workload as a tree of WorkloadItem nodes — a node with children compiles to a CompositePodGroupTemplate , a node without children to a PodGroupTemplate — and attaches its own defaults plus the user-supplied building blocks to each node. From there, Validate reports problems back at the exact field path within the controller's own API, BuildWorkload compiles the tree into a Workload, and NewPodGroup and NewCompositePodGroup stamp out the runtime group objects. Validation is deny-by-default: a controller declares the policies and disruption modes it actually supports through AllowedPolicies and AllowedDisruptionModes , and anything outside those lists is rejected. Building blocks added in future releases therefore stay unavailable until a controller explicitly opts into them. For hierarchical workloads where a parent controller owns the Workload and delegates group creation to its children, NewBuilderFromExistingWorkload lets a child materialize only its own PodGroup from the parent's Workload. Neither the building blocks nor the library have a feature gate of their own; they become user-visible through whichever controller adopts them. The native Job controller is the first to do so, and we detail it in the next section. Building upon the new controller integration APIs, the Job API now features an explicit .spec.scheduling field, so you can declare how a Job should be scheduled instead of relying on the Job controller to infer it from the Job's shape. This expands support well beyond static, indexed, and fully-parallel Jobs. .spec.scheduling is composed of the building blocks described above: schedulingPolicy — schedulingConstraints — the topology domain the Job's Pods must be co-located within. resourceClaims — the ResourceClaims shared by all of the Job's Pods. For example: apiVersion: batch/v1 kind: Job metadata: name: distributed-training-job annotations: kubernetes.io/description: "Distributed Job using explicit WAS scheduling with gang policy and zone topology constraints." spec: parallelism: 8 completions: 8 scheduling: schedulingPolicy: gang: {} minCount omitted → defaults to parallelism 8 schedulingConstraints: topology: - key: topology.kubernetes.io/zone disruptionMode: all: {} template: spec: containers: ... Omitting .spec.scheduling , or omitting schedulingPolicy within it, selects the basic policy, which behaves exactly like standard Job scheduling today. For every Job it manages, the controller compiles this configuration into a Workload and a PodGroup owned by the Job, and sets .spec.schedulingGroup.podGroupName on each Pod it creates so the scheduler treats them as one group. Once created, .spec.scheduling is immutable, with one exception: schedulingPolicy.gang.minCount can be updated, which lets you resize a running gang. As the core WAS APIs mature, so do their integrations with Dynamic Resource Allocation https://kubernetes.io/docs/concepts/resource-management/dynamic-resource-allocation/ DRA . Kubernetes v1.36 introduced the DRAWorkloadResourceClaims https://kubernetes.io/docs/reference/command-line-tools-reference/feature-gates/ DRAWorkloadResourceClaims feature gate. The associated feature allows ResourceClaims https://kubernetes.io/docs/concepts/resource-management/dynamic-resource-allocation/dra-api/ resourceclaims-templates to be replicated and reserved for entire PodGroups and shared by all their member Pods: apiVersion: scheduling.k8s.io/v1beta1 kind: PodGroup metadata: name: training-job-workers-pg spec: ... resourceClaims: - name: pg-claim resourceClaimTemplateName: my-claim-template --- apiVersion: v1 kind: Pod metadata: name: topology-aware-workers-pg-pod spec: ... schedulingGroup: podGroupName: training-job-workers-pg resourceClaims: - name: pg-claim resourceClaimTemplateName: my-claim-template In Kubernetes v1.37, the DRAWorkloadResourceClaims feature gate graduated to Beta. While the API and core functionality of the feature remain unchanged, one change eliminates some potentially surprising behavior when disabling the feature. Previously when one of a Pod's spec.resourceClaims referenced a ResourceClaimTemplate and matched one of its PodGroup's spec.resourceClaims and the DRAWorkloadResourceClaims feature gate was disabled , a ResourceClaim was created for the Pod instead of the PodGroup. In that scenario in v1.37, no ResourceClaim is created at all. This change prevents Kubernetes from creating a flood of ResourceClaims from a ResourceClaimTemplate and potentially exhausting DRA resources when a claim intended to be shared by a whole PodGroup is replicated for each and every Pod in the group. For more information, see the feature documentation https://kubernetes.io/docs/concepts/resource-management/dynamic-resource-allocation/dra-api/ workload-resource-claims . The Workload-Aware Scheduling Working Group https://www.kubernetes.dev/community/community-groups/wg/workload-aware-scheduling/ WG WAS is currently finalizing its plans for the Kubernetes v1.38 release cycle. While the roadmap is still taking shape stay tuned , the following key initiatives are already planned: Many of the workload-aware scheduling improvements are now available as Beta features in v1.37, while new advanced capabilities are introduced in Alpha. Both Beta and Alpha features here are disabled by default and require manual enablement. Beta features: GenericWorkload kube-apiserver , kube-controller-manager and kube-scheduler . Ensure your manifests are updated to use the scheduling.k8s.io/v1beta1 DRAWorkloadResourceClaims kubelet . Alpha features: Topology-aware scheduling: Enable the TopologyAwareWorkloadScheduling https://kubernetes.io/docs/reference/command-line-tools-reference/feature-gates/ TopologyAwareWorkloadScheduling feature gate on the kube-apiserver and kube-scheduler . CompositePodGroup API: Enable the CompositePodGroup https://kubernetes.io/docs/reference/command-line-tools-reference/feature-gates/ CompositePodGroup feature gate on the kube-apiserver , kube-controller-manager and kube-scheduler , and ensure the scheduling.k8s.io/v1alpha3 API version is enabled. Note that enabling CompositePodGroup on the kube-controller-manager also requires the TopologyAwareWorkloadScheduling https://kubernetes.io/docs/reference/command-line-tools-reference/feature-gates/ TopologyAwareWorkloadScheduling feature gate to be enabled. Workload API integration with the Job controller: Enable the WorkloadWithJob https://kubernetes.io/docs/reference/command-line-tools-reference/feature-gates/ WorkloadWithJob feature gate on the kube-apiserver and kube-controller-manager . PodGroup preemptionPolicy : Enable the PodGroupPreemptionPolicy https://kubernetes.io/docs/reference/command-line-tools-reference/feature-gates/ PodGroupPreemptionPolicy feature gate on the kube-apiserver and kube-scheduler . Controller integration APIs: The new workloadbuilder library is available to developers building both out-of-tree and in-tree controllers who want to integrate with WAS. It does not require a feature gate. You can explore the library and find usage examples directly in the kubernetes/component-helpers https://github.com/kubernetes/component-helpers/tree/master/scheduling/schedulingv1/workloadbuilder repository. We encourage you to try out workload-aware scheduling in your test clusters and share your experiences to help shape the future of Kubernetes scheduling. You can send your feedback by: To dive deeper into the architecture and design of these features, read the KEPs: