Workload

A workload is an application that will run to completion. It can be composed by one or multiple Pods that, loosely or tightly coupled, that, as a whole, complete a task. A workload is the unit of admission in Kueue.

The prototypical workload can be represented with a Kubernetes batch/v1.Job. For this reason, we sometimes use the word job to refer to any workload, and Job when we refer specifically to the Kubernetes API.

However, Kueue does not directly manipulate Job objects. Instead, Kueue manages Workload objects that represent the resource requirements of an arbitrary workload. Kueue automatically creates a Workload for each Job object and syncs the decisions and statuses.

The manifest for a Workload looks like the following:

apiVersion: kueue.x-k8s.io/v1beta1
kind: Workload
metadata:
  name: sample-job
  namespace: team-a
spec:
  active: true
  queueName: team-a-queue
  podSets:
  - count: 3
    name: main
    template:
      spec:
        containers:
        - image: gcr.io/k8s-staging-perf-tests/sleep:latest
          imagePullPolicy: Always
          name: container
          resources:
            requests:
              cpu: "1"
              memory: 200Mi
        restartPolicy: Never

Active

You can stop or resume a running workload by setting the Active field. The active field determines if a workload can be admitted into a queue or continue running, if already admitted. Changing .spec.Active from true to false will cause a running workload to be evicted and not be requeued.

Queue name

To indicate in which LocalQueue you want your Workload to be enqueued, set the name of the LocalQueue in the .spec.queueName field.

Pod sets

A Workload might be composed of multiple Pods with different pod specs.

Each item of the .spec.podSets list represents a set of homogeneous Pods and has the following fields:

• spec describes the pods using a v1/core.PodSpec.
• count is the number of pods that use the same spec.
• name is a human-readable identifier for the pod set. You can use the role of the Pods in the Workload, like driver, worker, parameter-server, etc.

Resource requests

Kueue uses the podSets resources requests to calculate the quota used by a Workload and decide if and when to admit a Workload.

Kueue calculates the total resources usage for a Workload as the sum of the resource requests for each podSet. The resource usage of a podSet is equal to the resource requests of the pod spec multiplied by the count.

Requests values adjustment

Depending on the cluster setup, Kueue will adjust the resource usage of a Workload based on:

• The cluster defines default values in Limit Ranges, the default values will be used if not provided in the spec.
• The created pods are subject of a Runtime Class Overhead.
• The spec defines only resource limits, case in which the limit values will be treated as requests.

Requests values validation

In cases when the cluster defines Limit Ranges, the values resulting from the adjustment above will be validated against the ranges. Kueue will mark the workload as Inadmissible if the range validation fails.

Reserved resource names

In addition to the usual resource naming restrictions, you cannot use the pods resource name in a Pod spec, as it is reserved for internal Kueue use. You can use the pods resource name in a ClusterQueue to set quotas on the maximum number of pods.

Priority

Workloads have a priority that influences the order in which they are admitted by a ClusterQueue. There are two ways to set the Workload priority:

• Pod Priority: You can see the priority of the Workload in the field .spec.priority. For a batch/v1.Job, Kueue sets the priority of the Workload based on the pod priority of the Job’s pod template.

• WorkloadPriority: Sometimes developers would like to control workload’s priority without affecting pod’s priority. By using WorkloadPriority, you can independently manage the priority of workloads for queuing and preemption, separate from pod’s priority.

Custom Workloads

As described previously, Kueue has built-in support for workloads created with the Job API. But any custom workload API can integrate with Kueue by creating a corresponding Workload object for it.

Dynamic Reclaim

It’s a mechanism allowing a currently Admitted workload to release a part of it’s Quota Reservation that is no longer needed.

Job integrations communicate this information by setting the reclaimablePods status field, enumerating the number of pods per podset for which the Quota Reservation is no longer needed.

status:
  reclaimablePods:
  - name: podset1
    count: 2
  - name: podset2
    count: 2
    

The count can only increase while the workload holds a Quota Reservation.

All or Nothing semantics for Job Resource Assignment

This mechanism allows a Job to be evicted and re-queued if the job doesn’t become ready. Please refer to the Sequential Admission with Ready Pods for more details.

Exponential Backoff Requeueing

Once evictions with PodsReadyTimeout reasons occur, a Workload will be re-queued with backoff. The Workload status allows you to know the following:

• .status.requeueState.count indicates the numbers of times a Workload has already been backoff re-queued by Eviction with PodsReadyTimeout reason
• .status.requeueState.requeueAt indicates the time when a Workload will be re-queued the next time

status:
  requeueState:
    count: 5
    requeueAt: 2024-02-11T04:51:03Z

When a Workload deactivated by Sequential Admission with Ready Pods is re-activated, the requeueState (.status.requeueState) will be reset to null.

What’s next

• Learn about workload priority class.
• Learn how to run jobs
• Read the API reference for Workload