scottk/kubernetes
1
0
Fork 0
mirror of https://github.com/optim-enterprises-bv/kubernetes.git synced 2026-01-07 15:51:45 +00:00
Code Issues Packages Projects Releases Wiki Activity

DRA scheduler: fix re-scheduling after ResourceSlice changes

Browse Source 
Making unschedulable pods schedulable again after ResourceSlice cluster events
was accidentally left out when adding structured parameters to Kubernetes 1.30.

All E2E tests were defined so that a driver starts first. A new test with a
different order (create pod first, wait for unschedulable, start driver)
triggered the bug and now passes.
This commit is contained in:
Patrick Ohly
2024-08-20 10:54:23 +02:00
parent 6dd2ade762
commit e85d3babf0
6 changed files with 212 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
pkg/scheduler/framework/plugins/dynamicresources/dynamicresources.go
View File

@@ -398,6 +398,8 @@ func (pl *dynamicResources) EventsToRegister(_ context.Context) ([]framework.Clu
{Event: framework.ClusterEvent{Resource: framework.Node, ActionType: framework.Add | framework.UpdateNodeLabel | framework.UpdateNodeTaint}},
// A pod might be waiting for a class to get created or modified.
{Event: framework.ClusterEvent{Resource: framework.DeviceClass, ActionType: framework.Add | framework.Update}},
// Adding or updating a ResourceSlice might make a pod schedulable because new resources became available.
{Event: framework.ClusterEvent{Resource: framework.ResourceSlice, ActionType: framework.Add | framework.Update}, QueueingHintFn: pl.isSchedulableAfterResourceSliceChange},
}
if pl.podSchedulingContextLister != nil {
@@ -491,6 +493,38 @@ func (pl *dynamicResources) isSchedulableAfterClaimChange(logger klog.Logger, po
return framework.Queue, nil
}
// isSchedulableAfterResourceSliceChange is invoked for add and update slice events reported by
// an informer. Such changes can make an unschedulable pod schedulable when the pod requests a device
// and the change adds a suitable device.
//
// For the sake of faster execution and avoiding code duplication, isSchedulableAfterResourceSliceChange
// only checks whether the pod uses claims. All of the more detailed checks are done in the scheduling
// attempt.
//
// The delete claim event will not invoke it, so newObj will never be nil.
func (pl *dynamicResources) isSchedulableAfterResourceSliceChange(logger klog.Logger, pod *v1.Pod, oldObj, newObj interface{}) (framework.QueueingHint, error) {
_, modifiedSlice, err := schedutil.As[*resourceapi.ResourceSlice](oldObj, newObj)
if err != nil {
// Shouldn't happen.
return framework.Queue, fmt.Errorf("unexpected object in isSchedulableAfterResourceSliceChange: %w", err)
}
if err := pl.foreachPodResourceClaim(pod, nil); err != nil {
// This is not an unexpected error: we know that
// foreachPodResourceClaim only returns errors for "not
// schedulable".
logger.V(6).Info("pod is not schedulable after resource slice change", "pod", klog.KObj(pod), "resourceSlice", klog.KObj(modifiedSlice), "reason", err.Error())
return framework.QueueSkip, nil
}
// We could check what got changed in the slice, but right now that's likely to be
// about the spec (there's no status yet...).
// We could check whether all claims use classic DRA, but that doesn't seem worth it.
// Let's assume that changing the slice may make the pod schedulable.
logger.V(5).Info("ResourceSlice change might make pod schedulable", "pod", klog.KObj(pod), "resourceSlice", klog.KObj(modifiedSlice))
return framework.Queue, nil
}
// isSchedulableAfterPodSchedulingContextChange is invoked for all
// PodSchedulingContext events reported by an informer. It checks whether that
// change made a previously unschedulable pod schedulable (updated) or a new