When a ResourceClaim status update hits a conflict, the retry replaces the whole status.devices list on the freshly fetched claim with a snapshot taken before the first attempt. Entries another driver wrote during the conflict window are discarded.
Where
pkg/driver/dra_hook.go, in PrepareResourceClaims:
// Store original devices list to preserve across conflict retries
originalDevices := claim.Status.Devices
err = wait.ExponentialBackoffWithContext(ctx, consts.Backoff, func(ctx context.Context) (bool, error) {
_, updateErr := d.client.ResourceV1().ResourceClaims(claim.Namespace).UpdateStatus(ctx, claim, metav1.UpdateOptions{})
if updateErr != nil {
if apierrors.IsConflict(updateErr) {
freshClaim, fetchErr := d.client.ResourceV1().ResourceClaims(claim.Namespace).Get(ctx, claim.Name, metav1.GetOptions{})
...
// Copy original devices list to fresh claim
freshClaim.Status.Devices = originalDevices
claim = freshClaimpkg/nri/nri.go has the same block in updateClaimNetworkDataWithRetry.
Why it matters
status.devices is shared by every driver that contributed a device to the claim, and each entry names its owner. Quoting k8s.io/api/resource/v1, AllocatedDeviceStatus:
Driver specifies the name of the DRA driver whose kubelet plugin should be invoked to process the allocation once the claim is needed on a node.
A conflict on this claim is most likely to come from exactly the other party that shares the list. Refetching and then overwriting the list with a pre-conflict copy undoes their write, and the retry then succeeds, so nothing reports it.
Suggested shape
Rebuild this driver's entries on top of the latest claim on every retry rather than restoring a whole-list snapshot: keep every entry whose Driver is not this driver, and replace or append this driver's own. AllocatedDeviceStatus identifies a device by driver, pool, device and share ID, so the merge key needs all four rather than the device name alone.
I have this implemented and can send a PR referencing this issue.
When a ResourceClaim status update hits a conflict, the retry replaces the whole
status.deviceslist on the freshly fetched claim with a snapshot taken before the first attempt. Entries another driver wrote during the conflict window are discarded.Where
pkg/driver/dra_hook.go, inPrepareResourceClaims:pkg/nri/nri.gohas the same block inupdateClaimNetworkDataWithRetry.Why it matters
status.devicesis shared by every driver that contributed a device to the claim, and each entry names its owner. Quotingk8s.io/api/resource/v1,AllocatedDeviceStatus:A conflict on this claim is most likely to come from exactly the other party that shares the list. Refetching and then overwriting the list with a pre-conflict copy undoes their write, and the retry then succeeds, so nothing reports it.
Suggested shape
Rebuild this driver's entries on top of the latest claim on every retry rather than restoring a whole-list snapshot: keep every entry whose
Driveris not this driver, and replace or append this driver's own.AllocatedDeviceStatusidentifies a device by driver, pool, device and share ID, so the merge key needs all four rather than the device name alone.I have this implemented and can send a PR referencing this issue.