Built-in integrations, grouped by common use:

• Batch and training: Kubernetes Job (including CronJobs), JobSet, MPIJob, RayJob (KubeRay), TrainJob (Kubeflow Trainer v2) and SparkApplication (alpha).
• Inference and serving: LeaderWorkerSet, RayService, Deployment and StatefulSet.
• Other workloads: Pods and Pod groups, AppWrapper and RayCluster.

External and custom integrations: run external frameworks, such as Argo Workflows and Tekton, or write an integration for your own workload types.

Deprecated: Kubeflow Trainer v1 jobs (PyTorchJob, TFJob, JAXJob, PaddleJob, XGBoostJob); migrate to TrainJob or JobSet.

Priorities include integrating with Workload-Aware Scheduling (WAS) in kube-scheduler and continued improvements for training, inference and other workloads.

2026 priorities

• Integrate with Workload-Aware Scheduling (WAS) in kube-scheduler, including Topology-Aware Scheduling #8871
• MultiKueue improvements:
  • Extend support for elastic Ray workloads, including autoscaling #13243 (replaces #8712)
  • Add workload-level admission constraints and preference-aware dispatching (proposal #8729)
  • Prevent starting preemptions in multiple worker clusters #8303
  • Maintain service availability during cross-cluster moves (proposal #8526)
  • Retrieve logs from worker clusters #3526
• Topology-Aware Scheduling improvements:
  • Support resource transformations #8860
  • Extend support for elastic workloads #8160
  • Evict workloads from nodes that become tainted #8828
• Support concurrent workload admission #8691
• Support large, high-priority jobs ("hero workloads") #8826
• Consider preemption cost when finding preemption candidates #7990
• Extend the Dynamic Resource Allocation (DRA) integration #8243, #2941

Long-term goals

• Partial preemption of serving workloads #3762
• Deeper integration with workflow frameworks such as Argo Workflows and Tekton #74
• Budget support #28
• Flavor assignment strategies, e.g. minimizing cost vs minimizing borrowing #312
• Cooperative preemption for workloads that implement checkpointing #477
• Delayed preemption for two-stage admission #3758
• Graduate the API to v1 #3476

• Understand queueing: workloads, queueing and admission.
• Configure queues: administer cluster quotas, then set priorities and queueing strategies.
• Share quota: cohorts, borrowing, fair-sharing policies, preemption, flavor fungibility and dynamic reclaim.
• Size jobs: start a Job smaller with partial admission, or resize supported jobs with elastic workloads.
• Control admission: admission checks, autoscaling with ProvisioningRequest, all-or-nothing scheduling, Pod readiness and enforcing Kueue management.
• Place workloads: set up Topology-Aware Scheduling and MultiKueue.
• Manage device quotas: DRA concepts and DRA setup; consumable capacity is alpha.
• Run in production: production setup (cert-manager, network policies, Prometheus) and RBAC.
• Monitor Kueue: observability, metrics, pending workloads and the KueueViz dashboard.

Browse concepts, task guides, examples and DeepWiki.

API stability: core APIs use v1beta2 and follow the Kubernetes deprecation policy.

Testing: the dashboards below track the main branch, and end-to-end suites run on Kind clusters. For supported Kubernetes versions, see the installation prerequisites; for scalability, see the performance test setup.