#Configuring preemption

Preemption is the process of evicting admitted workloads so that higher-priority or fairly-shared workloads can be admitted. Alauda Build of Kueue supports several preemption strategies that allow administrators to control how workloads compete for resources.

#Preemption policies

Alauda Build of Kueue supports the following global preemption policies, which can be configured when deploying the Alauda Build of Kueue cluster plugin:

#ClusterQueue preemption configuration

You can configure preemption behavior at the ClusterQueue level using the spec.preemption field. This allows fine-grained control over when and how workloads in a specific ClusterQueue can be preempted.

#Example ClusterQueue with preemption configuration

apiVersion: kueue.x-k8s.io/v1beta2
kind: ClusterQueue
metadata:
  name: team-a-queue
spec:
  namespaceSelector: {}
  resourceGroups:
  - coveredResources: ["cpu", "memory", "nvidia.com/gpu"]
    flavors:
    - name: "default-flavor"
      resources:
      - name: "cpu"
        nominalQuota: 16
      - name: "memory"
        nominalQuota: 64Gi
      - name: "nvidia.com/gpu"
        nominalQuota: 4
  preemption:
    reclaimWithinCohort: Any
    borrowWithinCohort:
      policy: LowerPriority
    withinClusterQueue: LowerPriority

#Preemption field reference

#withinClusterQueue

Controls whether a pending workload can preempt other workloads in the same ClusterQueue.

#reclaimWithinCohort

Controls whether a pending workload can preempt workloads in other ClusterQueues of the same cohort to reclaim nominal quota that is being borrowed.

#borrowWithinCohort

Controls whether a pending workload can preempt workloads in other ClusterQueues of the same cohort to borrow unused resources.

#Non-preemptive queues

Some workload types, such as interactive notebook sessions, are not suspendable. These workloads should only be assigned to a local queue backed by a non-preemptive ClusterQueue.

A non-preemptive ClusterQueue keeps the default preemption settings (all set to Never):

apiVersion: kueue.x-k8s.io/v1beta2
kind: ClusterQueue
metadata:
  name: non-preemptive-queue
spec:
  namespaceSelector: {}
  resourceGroups:
  - coveredResources: ["cpu", "memory"]
    flavors:
    - name: "default-flavor"
      resources:
      - name: "cpu"
        nominalQuota: 16
      - name: "memory"
        nominalQuota: 64Gi
  preemption:
    withinClusterQueue: Never
    reclaimWithinCohort: Never
    borrowWithinCohort:
      policy: Never