Overview

Kueue is a kubernetes-native system that manages quotas and how jobs consume them. Kueue decides when a job should wait, when a job should be admitted to start (as in pods can be created) and when a job should be preempted (as in active pods should be deleted).

Why use Kueue

You can install Kueue on top of a vanilla Kubernetes cluster. Kueue does not replace any existing Kubernetes components. Kueue is compatible with cloud environments where:

• Compute resources are elastic and can be scaled up and down.
• Compute resources are heterogeneous (in architecture, availability, price, etc.).

Kueue APIs allow you to express:

• Quotas and policies for fair sharing among tenants.
• Resource fungibility: if a resource flavor is fully utilized, Kueue can admit the job using a different flavor.

A core design principle for Kueue is to avoid duplicating mature functionality in Kubernetes components and well-established third-party controllers. Autoscaling, pod-to-node scheduling and job lifecycle management are the responsibility of cluster-autoscaler, kube-scheduler and kube-controller-manager, respectively. Advanced admission control can be delegated to controllers such as gatekeeper.

Features overview

• Job management: Support job queueing based on priorities with different strategies: StrictFIFO and BestEffortFIFO.
• Advanced Resource management: Comprising: resource flavor fungibility, fair sharing, cohorts and preemption with a variety of policies between different tenants.
• Integrations: Built-in support for popular jobs, e.g. BatchJob, Kubeflow training jobs, RayJob, RayCluster, JobSet, plain Pod and Pod Groups.
• System insight: Build-in prometheus metrics to help monitor the state of the system, and on-demand visibility endpoint for monitoring of pending workloads.
• AdmissionChecks: A mechanism for internal or external components to influence whether a workload can be admitted.
• Advanced autoscaling support: Integration with cluster-autoscaler’s provisioningRequest via admissionChecks.
• All-or-nothing with ready Pods: A timeout-based implementation of All-or-nothing scheduling.
• Partial admission and dynamic reclaim: mechanisms to run a job with reduced parallelism, based on available quota, and to release the quota the pods complete..
• Mixing training and inference: Simultaneous management of batch workloads along with serving workloads (such as Deployments or StatefulSets)
• Multi-cluster job dispatching: called MultiKueue, allows to search for capacity and off-load the main cluster.
• Topology-Aware Scheduling: Allows to optimize the pod-pod communication throughput by scheduling aware of the data-center topology.

Job-integrated features

High-level Kueue operation

To learn more about Kueue concepts, see the concepts section.

To learn about different Kueue personas and what you can do with Kueue, see the tasks section.