Catalog / Kubernetes

1.36.3

3 months agoaddedfixedchangedOriginal notes

Changelog since v1.36.2

Changes by Kind

Feature

Failing Test

  • Fixed a bug when the DRADeviceTaintRules feature is enabled that caused kube-scheduler to panic when DeviceTaintRules exist and ResourceSlices are changed or to ignore new changes to DeviceTaintRules. (#139681, @nojnhuh) [SIG Node and Testing]

Bug or Regression

  • Cri-api: Reverts to pre-1.34 JSON encoding of the KeyValue value field (#139965, @liggitt) [SIG Node]
  • Fixed DRA scheduling bugs where the structured allocator mis-counted a device's shared counters while exploring candidates: it could keep a counter reserved after rejecting or backtracking a candidate, or drop a shared device's in-use marker so a later share was charged the counter twice. Either way the allocator could treat a counter set as exhausted and leave a pod pending on a node that could satisfy it. This affected the allocator used by the default feature configuration. (#140663, @thc1006) [SIG Node]
  • Fixed a kubelet memory leak regression in 1.36 caused by leaked contexts on every Pod sync. (#140066, @compumike) [SIG Node]
  • Fixes a 1.36 regression in server side apply where patching a container type (list or map) could result in 422 required errors for apply requests that previously succeeded. (#140296, @jpbetz) [SIG API Machinery, Architecture, Auth, CLI, Cloud Provider, Cluster Lifecycle, Network, Node, Scheduling and Storage]
  • Kubeadm: Improved resilience of kubeadm etcd learner promotion. kubeadm now correctly handles cases where learner promotion succeeds on the etcd side but a transient client-side error is returned, preventing unnecessary etcd-join failures. (#139910, @jihyun-huh) [SIG Cluster Lifecycle]
  • Kubeadm: during "kubeadm join", use the KubernetesAPICall timeout (default 1 minute) when fetching the kubeadm-config ConfigMap, instead of the short 350ms retry used for optional component configs. A new shortConfigMapGet parameter is added to FetchInitConfigurationFromCluster so that callers like "kubeadm reset" can still use the short retry. (#139808, @damdo) [SIG Cluster Lifecycle]
  • Kubeadm: fix MemberPromote to skip the etcd promote API call when the member is already a voting member, avoiding unnecessary retries and timeout. (#138493, @wgkingk) [SIG Cluster Lifecycle]
  • Kubelet no longer emits V(4) "Label not found" logs for missing optional container annotations. (#140322, @HirazawaUi) [SIG Node]

Dependencies

Added

Nothing has changed.

Changed

Removed

Nothing has changed.