Skip to main content

LeaderWorkerSet Support

When the LeaderWorkerSet CRD (leaderworkerset.x-k8s.io/v1) is installed, PDB Operator also manages PDBs for LWS workloads such as multi-host vLLM or SGLang deployments. Support is detected at startup; without the CRD the operator runs unchanged.

Added in operator v0.4.0.

Why groups, not pods​

An LWS group restarts as a unit. Under the default RecreateGroupOnPodRestart, evicting a single pod takes down all size pods in its group.

A budget that counts individual pods is wrong in both directions:

  • It under-protects capacity, because losing one pod really means losing a whole group
  • It can deadlock node drains, because the percentage never lines up with a group boundary

So the operator quantizes the budget to whole groups.

The calculation​

desiredGroups = ceil(class% × replicas), clamped to replicas - 1
minAvailable  = desiredGroups × size

Worked example, replicas: 4 and size: 8 under mission-critical (90%):

StepValue
ceil(0.90 × 4)4
clamped to replicas - 13
minAvailable = 3 × 824

Exactly one group may be disrupted at a time, and a drain always makes progress.

Granularity is whole groups, so a 4-group set has only 4 protection steps. Availability classes that land between group boundaries round to the nearest whole group.

Special cases​

ShapeBehavior
replicas: 1No PDB is created. Any budget over a single group would permanently block node drains, so the operator emits a Warning event explaining why instead
size: 1Plain pod-level semantics, identical to a Deployment
custom with absolute minAvailableRounded up to the next whole group, so a hand-set budget can never split one
LWS CRD absentThe controller is not registered; nothing else changes

Pod selection and StatefulSet overlap​

The managed PDB selects on the leaderworkerset.sigs.k8s.io/name label, which covers both leader and worker pods.

LWS implements each set as a leader StatefulSet plus per-group worker StatefulSets. The StatefulSetController skips those (detected by the same label) so a pod never matches more than one managed PDB. If it did, the eviction API would reject every eviction for those pods.

Example​

apiVersion: availability.pdboperator.io/v1alpha1
kind: PDBPolicy
metadata:
  name: inference-critical
spec:
  availabilityClass: mission-critical
  enforcement: strict
  workloadSelector:
    matchLabels:
      workload: inference

Applied to a LeaderWorkerSet with replicas: 4 and size: 8, this produces a PDB with minAvailable: 24.

Verify what was created:

kubectl get pdb -l pdboperator.io/managed-by=pdb-operator
kubectl get events --field-selector involvedObject.name=<lws-name>

Metrics​

pdb_operator_leaderworkersets_managed reports currently managed LeaderWorkerSets by namespace and availability class. See the Metrics Reference.