TL;DR — Kueue is a Kubernetes-native job queueing system. It doesn't place pods (the scheduler does that) — it decides when a whole job is allowed to start, based on quotas and fairness across teams. ClusterQueue defines a resource pool, LocalQueue is a team's entry point, and cohorts let queues borrow each other's idle GPUs. It's an official Kubernetes SIG project.
What it is
Kueue is a job-queueing and quota-management layer that sits above the Kubernetes scheduler. A job submitted to Kueue waits in a queue until its quota is available; only then does Kueue admit it and let the normal scheduler place the pods. It's maintained by kubernetes-sigs — as close to "official" as add-ons get. In the AI Native landscape it's in AI Native Infra › Orchestration and Scheduling.
Why it exists
On a shared GPU cluster the real question isn't "where does this pod go" but "whose job runs now, and how much can each team consume." Without that control, whoever submits first grabs all the GPUs and everyone else starves. Kueue adds the missing governance: per-team quotas, fair sharing, priorities, and borrowing — so an expensive cluster stays both fully used and fairly used.
How it works — the objects
| Object | Role |
|---|---|
ResourceFlavor | A flavor of hardware — e.g. A100 vs H100 vs spot — so quota can target specific accelerators. |
ClusterQueue | Cluster-scoped resource pool: how much CPU/mem/GPU of each flavor is available, plus policies. |
LocalQueue | Namespaced pointer a team submits to; routes their jobs to a ClusterQueue. |
Cohort | A group of ClusterQueues that can borrow one another's unused quota. |
Workload | Kueue's internal record of a queued job awaiting admission. |
Fig 1 — Teams submit to LocalQueues; Kueue admits jobs by quota, then the scheduler places them.
Key features
- Quotas & fair sharing — per-team caps with weighted fairness across the cohort.
- Borrowing & preemption — idle quota is lent to busy queues, then reclaimed (preempting borrowed jobs) when the owner needs it.
- Partial admission & dynamic reclaim — run a job at reduced parallelism if full quota isn't free; release quota as pods finish.
- ResourceFlavor fungibility — fall back across hardware types (e.g. spill to spot or a different GPU).
- Broad job support — BatchJob, Kubeflow training jobs, RayJob, RayCluster, JobSet, plain Pods and Pod groups.
- MultiKueue — dispatch jobs across multiple clusters (a 2026 focus area).
Quick start
Install the controller, then create one ClusterQueue + a LocalQueue and submit jobs against it:
kubectl apply --server-side -f https://github.com/kubernetes-sigs/kueue/releases/latest/download/manifests.yaml
kubectl get pods -n kueue-system # controller up
kubectl get clusterqueues # after you create oneA job opts in by labeling itself with its LocalQueue; Kueue suspends it until quota frees up, then releases it to the scheduler:
metadata:
labels:
kueue.x-k8s.io/queue-name: team-a-lqWhen to use, when to skip
Use it when multiple teams share a GPU/TPU cluster and you need quotas, fairness, and borrowing — the classic "platform team rations scarce accelerators" problem. It layers cleanly on the existing scheduler, so adoption is low-risk.
Pair, don't pick. Kueue handles admission; it doesn't gang-schedule placement. For distributed training you often run Kueue with Volcano (Kueue for quota, Volcano for gang). For a single-tenant cluster with no quota needs, you may not need Kueue at all.
vs the alternatives
| Tool | Best for | Trade-off |
|---|---|---|
| Kueue | Job queueing, quota, fair-share, borrowing | Admission only — not a placement scheduler |
| Volcano | Gang scheduling + rich placement policies | A full scheduler to run |
| YuniKorn | Unified batch+service scheduling with queues | Replaces the scheduler |
References
- Official documentation — overview & tutorials.
- Concepts — queues, flavors, cohorts.
- ClusterQueue — quota model in detail.
- kubernetes-sigs/kueue — source.
Extra reads
- Quota sharing across namespaces on GKE — cohort tutorial.
- Fair-share GPU/TPU queueing — hands-on quick start.
- Batch scheduling on K8s compared — Kueue vs Volcano vs YuniKorn.
Verified against the official Kueue docs (kueue.sigs.k8s.io), May 2026.