Kueue: Kubernetes-native Job Queueing
Manage quotas and resource sharing for batch, HPC, and AI/ML workloads. Kueue decides when and where jobs should run, integrating seamlessly with your existing ecosystem.
Core Capabilities
Kueue governs resource admission natively for standard APIs (including Job, JobSet, KubeRay, Kubeflow Trainer, and LeaderWorkerSet) using these five core architectural pillars:
Multi-Tenant Quotas
Define resource quotas globally at the cluster level and share them securely across different teams and namespaces. Govern resource allocations, queue priorities, and limits using native Kubernetes APIs.
Dynamic Capacity Sharing
Group queues together to enable dynamic resource sharing. Workloads can automatically borrow unused capacity from peer teams when their resource pools are idle, maximizing overall GPU/CPU utilization and slashing costs.
Topology-Aware Placement
Optimize the placement of large, distributed machine learning training workloads (like LLMs) based on physical network topology. Minimize network latency by grouping tightly-coupled pods on the same infrastructure.
Resource Pools & Preemption
Support diverse hardware pools, such as Spot vs. On-Demand instances or different GPU models. Workloads are automatically routed to the appropriate resource pool, with priority-based preemption ensuring higher-priority jobs run first.
Multi-Cluster Scaling
Scale workloads seamlessly beyond a single cluster. A central control plane automatically dispatches jobs to multiple worker clusters based on global resource availability, enabling hybrid and multi-cloud scheduling.
Architecture & Integration
Kueue acts as an admission controller before the scheduler, managing queues and quotas cleanly without replacing core Kubernetes components.
Join the Conversation
Kueue is an open, community-driven CNCF project. Get involved!