Kueue

Kueue is a free, open-source Kubernetes project for managing quotas and job resource use. It determines when workloads wait, start or are preempted, and installs on a standard Kubernetes cluster without replacing existing components. Scheduling options include priority-based queues, fair sharing, quota borrowing between teams, resource flavor fungibility, preemption policies and all-or-nothing scheduling. Built-in integrations cover Kubernetes Jobs and Pods, Kubeflow training jobs, Ray jobs and clusters, JobSet, AppWrapper, Deployments, StatefulSets and LeaderWorkerSet. MultiKueue can find capacity and dispatch work across clusters. Topology-aware scheduling considers data-center layout to improve communication throughput between pods. For visibility, Kueue provides Prometheus metrics and an endpoint for pending workloads. It installs batch administrator and user ClusterRoles, with resource-specific roles that can be combined into custom roles. Its webhook server includes internal certificate management, and installation guidance also covers cert-manager. The installation guide recommends Kubernetes 1.34 or newer; only those versions are officially supported and tested. Installation options include kubectl, Helm or a source build.

Kueue provides Kubernetes workload queueing, quota controls and scheduling features, including multi-cluster dispatch. Teams should note the documented Kubernetes version requirement before installing.

Kueue suits teams running batch or other supported workloads on Kubernetes that need quota management and scheduling controls. It is designed for administrators managing shared cluster resources across teams or clusters.

Kueue is a Kubernetes-native control plane for deciding when batch workloads consume shared resources. It is best suited to teams already operating Kubernetes clusters and needing enforceable quotas, fair access, and queue policies across varied job types. The free, self-hosted model avoids license cost, but its officially supported Kubernetes baseline makes platform version a key buying decision.

For Kubernetes operators, Kueue adds sophisticated workload governance without replacing the cluster components already in place. Teams running another scheduler or without Kubernetes should look elsewhere.

Overview

Kueue manages job resource consumption through policies that determine whether workloads wait, start, or are preempted. It installs on a vanilla Kubernetes cluster, so adoption can extend an existing environment rather than introduce a replacement cluster manager. That makes it a practical layer for shared batch infrastructure, while its Kubernetes dependency rules out standalone use on conventional servers or unrelated orchestration platforms.

The project is an open, community-driven CNCF project with Slack, GitHub, and mailing-list channels. Installation guidance covers kubectl, Helm, source builds, and development versions. The installation guide recommends Kubernetes 1.34 or newer, and only those versions are officially supported and tested; clusters on older releases face a clear compatibility risk.

Key features

Queues, quotas, and sharing

Priority-based queueing lets administrators rank work instead of relying solely on submission order. Tenant quotas, cohorts, fair sharing, shared resource pools, and quota borrowing provide mechanisms for dividing capacity between teams while allowing unused allocation to be used elsewhere. Resource flavor fungibility can keep jobs moving when equivalent capacity exists in another flavor.

Preemption policies determine when running work can yield to higher-priority demand. All-or-nothing scheduling is useful for distributed jobs that must receive their requested resources together, rather than starting partially and waiting indefinitely. These controls are powerful for multi-tenant clusters, but they also demand deliberate policy design and ongoing Kubernetes administration.

Workload integrations and placement

Built-in integrations cover Kubernetes Batch Jobs, plain Pods and Pod groups, Deployments, StatefulSets, JobSet, AppWrapper, Kubeflow training jobs, Ray jobs and clusters, and LeaderWorkerSet. That breadth reduces the need to create separate scheduling paths for common AI, data-processing, and distributed workloads.

MultiKueue can search for capacity and dispatch jobs across clusters, extending queue policy beyond a single installation. Topology-aware scheduling places workloads with data-center topology in mind to improve pod-to-pod communication throughput, a meaningful advantage for tightly coupled distributed jobs.

Operations and access

Prometheus metrics support ongoing monitoring, while an on-demand visibility endpoint exposes pending workloads when operators need a queue view. Kueue installs batch administrator and batch user ClusterRoles and allows custom roles to be composed from resource-specific editor and viewer roles. Its webhook server uses internal certificate management, with documentation also covering cert-manager for teams that prefer that approach.

Pricing

Kueue costs nothing under its open-source Kubernetes project plan. There are no paid tiers, seat charges, trial periods, or renewal terms to budget for. The trade-off is operational: teams provide their own Kubernetes infrastructure, upgrades, policy administration, monitoring, and support processes rather than receiving a hosted service.

This plan fits organizations that already run Kubernetes and want queueing and quota controls without a software license. It is less suitable for a small team seeking a managed scheduler with vendor-operated infrastructure or a turnkey user interface.

Platforms

Kueue is delivered as an API-oriented, self-hosted deployment on Kubernetes. Installation can use kubectl or Helm, or teams can build from source. The platform requirement is specific: Kubernetes 1.34 or newer is the officially supported and tested baseline.

Who it's for

Kueue fits platform and batch-infrastructure teams sharing GPU, CPU, or other Kubernetes resources among tenants and job classes. It is particularly relevant for organizations running Kubeflow, Ray, JobSet, or other supported workload forms that need consistent queue, quota, and preemption behavior. Multi-cluster operators and teams with communication-sensitive distributed jobs also gain from MultiKueue and topology-aware placement.

It is not the right choice for teams without Kubernetes, for users who need a hosted product rather than self-hosted operations, or for clusters that cannot meet the documented Kubernetes version requirement.

Pros and cons

Pros

• Deep capacity policy: quotas, fair sharing, cohorts, borrowing, fungibility, priority, and preemption address complex multi-tenant scheduling needs.
• Broad workload coverage: integrations span core Kubernetes objects, AI-training systems, Ray, and distributed workload controllers.
• Cluster-aware dispatch: MultiKueue can send jobs to clusters with available capacity, while topology-aware scheduling targets better communication efficiency.
• No license cost: the open-source plan is free and self-hosted.
• Operator tooling: Prometheus metrics, pending-workload visibility, role composition, and certificate options support controlled administration.

Cons

• Kubernetes-only operation: teams without Kubernetes cannot use it as their scheduler.
• Version constraint: official support and testing begin at Kubernetes 1.34, which can block older clusters.
• Self-hosted responsibility: infrastructure, upgrades, security configuration, and day-to-day policy work remain with the adopter.
• Policy complexity: advanced sharing and preemption controls reward experienced Kubernetes administrators and require careful governance.

Alternatives

Backend.AI is a better fit when you want a freemium platform with web access, Linux support, self-hosting, and a home-user open-source plan alongside a trial option.

Choose dstack when you need an orchestration stack spanning Windows, macOS, Linux, web, and self-hosted deployments, with an optional GPU marketplace.

HAMi is the narrower alternative for open-source GPU virtualization middleware on Linux and self-hosted Kubernetes, without a free trial.

Pick NVIDIA ShadowPlay for Windows game capture on a supported GeForce system, not cluster workload scheduling.

ClearML suits teams wanting a freemium, cross-platform machine-learning platform with an open-source self-hosted version and a no-credit-card community plan.

GPUStack is a straightforward open-source GPU cluster manager with broad desktop and server platform coverage.

TORQUE Resource Manager is the alternative for a traditional open-source resource manager across API and self-hosted Linux environments.

Apache YuniKorn is another free, self-hosted scheduler when an Apache-licensed project and a source-build installation model fit your operation.

Verdict

Choose Kueue if your organization already runs Kubernetes and needs rigorous, multi-tenant batch scheduling with quotas, fair sharing, preemption, and integrations across modern workload controllers. Its main reason to choose it is policy depth without license fees or a replacement cluster manager. Look elsewhere if you need a hosted service, a scheduler independent of Kubernetes, or support for a cluster older than the officially supported baseline.