Gang scheduling

This quickstart creates two LeaderWorkerSet replicas of two pods each. The empty spec.scheduling: {} opts into native Workload-Aware Scheduling and selects the default behavior: the leader and worker in each replica form an independent gang. Kubernetes admits both members of a replica together or keeps both pending.

Before starting, complete the native gang scheduling installation prerequisites.

Create the LeaderWorkerSet

The example requests only CPU and memory, so it does not require accelerators or a device plugin.

apiVersion: leaderworkerset.x-k8s.io/v1
kind: LeaderWorkerSet
metadata:
  name: lws-gang
spec:
  replicas: 2
  scheduling: {}
  leaderWorkerTemplate:
    size: 2
    workerTemplate:
      spec:
        containers:
        - name: nginx
          image: nginxinc/nginx-unprivileged:1.27
          resources:
            requests:
              cpu: 100m
              memory: 32Mi

Apply it from the documentation site:

kubectl apply -f https://lws.sigs.k8s.io/examples/leaderworkerset/gang-scheduling/lws.yaml

Wait for both replicas to become ready:

kubectl wait leaderworkerset/lws-gang \
  --for=jsonpath='{.status.readyReplicas}'=2 --timeout=5m

Inspect the scheduling objects

LWS creates one Workload and two PodGroup objects. Each PodGroup has a gang minimum of two, matching the replica size:

kubectl get workloads,podgroups \
  -l leaderworkerset.sigs.k8s.io/name=lws-gang

kubectl get podgroups \
  -l leaderworkerset.sigs.k8s.io/name=lws-gang \
  -o custom-columns='NAME:.metadata.name,MIN-COUNT:.spec.schedulingPolicy.gang.minCount,INITIALLY-SCHEDULED:.status.conditions[?(@.type=="PodGroupInitiallyScheduled")].status'

Confirm that every pod names its PodGroup and has been assigned a node:

kubectl get pods \
  -l leaderworkerset.sigs.k8s.io/name=lws-gang \
  -o custom-columns='NAME:.metadata.name,NODE:.spec.nodeName,PODGROUP:.spec.schedulingGroup.podGroupName'

For the scheduling lifecycle and restrictions, see the gang scheduling concepts.

Clean up

Deleting the LeaderWorkerSet also deletes its Workload and PodGroups:

kubectl delete -f https://lws.sigs.k8s.io/examples/leaderworkerset/gang-scheduling/lws.yaml