- `requests.deviceclass.resource.k8s.io/<deviceclass>` is charged based on the worst-case number of devices requested.
  - Device classes mapping to an extended resource now consume `requests.<extended resource name>`. ([#134210](https://github.com/kubernetes/kubernetes/pull/134210), [@yliaog](https://github.com/yliaog)) [SIG API Machinery, Apps, Node, Scheduling and Testing]

### SIG Cloud Provider

  final long expireAfterWriteNanos;

  /** How long after the last write an entry becomes a candidate for refresh. */
  final long refreshNanos;

  /** Entries waiting to be consumed by the removal listener. */
  // TODO(fry): define a new type which creates event objects and automates the clear logic
  final Queue<RemovalNotification<K, V>> removalNotificationQueue;

  /**

- When `SchedulerQueueingHints` is enabled, clear events cached in the scheduling queue as soon as possible so that the scheduler consumes less memory. ([#120586](https://github.com/kubernetes/kubernetes/pull/120586), [@sanposhiho](https://github.com/sanposhiho)) [SIG Scheduling]

- Fixes a flaw (CVE-2019-11253) in json/yaml decoding where large or malformed documents could consume excessive server resources. Request bodies for normal API requests (create/delete/update/patch operations of regular resources) are now limited to 3MB. ([#83261](https://github.com/kubernetes/kubernetes/pull/83261), [@liggitt](https://github.com/liggitt))...

		* Restored the ability of kubectl running inside a pod to consume resource files specifying a different namespace than the one the pod is running in. ([#44862](https://github.com/kubernetes/kubernetes/pull/44862), [@liggitt](https://github.com/liggitt))

* Update the GCE windows node image to include hot fixes since July. ([#81106](https://github.com/kubernetes/kubernetes/pull/81106), [@YangLu1031](https://github.com/YangLu1031))

* Fix aggressive VM calls for Azure VMSS ([#83102](https://github.com/kubernetes/kubernetes/pull/83102), [@feiskyer](https://github.com/feiskyer))

- Dynamic resource allocation: when a claim uses "wait for first consumer" allocation (the default), then it will now get deallocated after it was used by a pod. That ensures that the next pod isn't affected by previous scheduling decision and that resources are not kept allocated unless really needed. If keeping a claim...

  - example manifests for the kube-apiserver, kube-controller-manager, kube-schedule, kubelet, and the cloud-controller-manager
  - labels/annotations that are consumed by any cloud specific controllers

In addition to the above, SIG Cloud Provider has been focusing on a long running effort to remove cloud provider code from kubernetes/kubernetes.

### SIG-cluster-lifecycle