* This is a breaking change (and only slightly) if you used dependencies with `yield`, used `except` in those dependencies, and didn't raise again.

`DeviceRequestAllocationResult` struct now has an "AdminAccess" field which should be used instead of the corresponding field in the `DeviceRequest` field when dealing with an allocation. If a device is only allocated for admin access, allocating it again for normal usage is now supported, as originally intended. To allow admin access, starting with 1.32 the `DRAAdminAccess` feature gate must be enabled. ([#127266](https://github.com/kubernetes/kubernetes/pull/127266), [@pohly](https://github.com/pohly))...

		default:
			return re, fmt.Errorf("replication mrf: unknown version: %d", binary.LittleEndian.Uint16(data[2:4]))
		}

		// OK, parse data.
		// ignore any parsing errors, we do not care this file is generated again anyways.
		re.DecodeMsg(msgp.NewReader(rc))

		return re, nil
	}

	globalLocalDrivesMu.RLock()
	localDrives := cloneDrives(globalLocalDrivesMap)
	globalLocalDrivesMu.RUnlock()

				storeFilePath := pathJoin(deletePath, ".DS_Store")
				_, err := Stat(storeFilePath)
				// .DS_Store exists
				if err == nil {
					// delete first
					Remove(storeFilePath)
					// try again
					Remove(deletePath)
				}
			}
			// Ignore errors if the directory is not empty. The server relies on
			// this functionality, and sometimes uses recursion that should not

- After DBus restarts, make `GracefulNodeShutdown` work again ([#100369](https://github.com/kubernetes/kubernetes/pull/100369), [@wzshiming](https://github.com/wzshiming))

- Fix: After a Node is down and take some time to get back to up again, the mount point of the evicted Pods cannot be cleaned up successfully. (#111933) Meanwhile Kubelet will print the log `Orphaned pod "xxx" found, but error not a directory occurred when trying to remove the volumes dir` every 2 seconds. (#105536) ...

- Correct a reference to a not/no longer used kustomize subcommand in the documentation ([#82535](https://github.com/kubernetes/kubernetes/pull/82535), [@demobox](https://github.com/demobox))
- kubectl could scale custom resource again ([#81342](https://github.com/kubernetes/kubernetes/pull/81342), [@knight42](https://github.com/knight42))

## Changes by Kind

### API Change

- Make STS available replicas optional again, ([#109241](https://github.com/kubernetes/kubernetes/pull/109241), [@ravisantoshgudimetla](https://github.com/ravisantoshgudimetla)) [SIG API Machinery and Apps]
- Make STS available replicas optional again. ([#113122](https://github.com/kubernetes/kubernetes/pull/113122), [@ashrayjain](https://github.com/ashrayjain)) [SIG Apps]

* GCE bring-up: Differentiate NODE_TAGS from NODE_INSTANCE_PREFIX ([#29141](https://github.com/kubernetes/kubernetes/pull/29141), [@zmerlynn](https://github.com/zmerlynn))
* hyperkube: fix build for 3rd party registry (again) ([#28489](https://github.com/kubernetes/kubernetes/pull/28489), [@liyimeng](https://github.com/liyimeng))

API server now exits if it encounters a networking failure (e.g. the networking interface hosting its address goes away) to allow a process manager (systemd/kubelet/etc) to react to the problem. Previously the server would log the failure and try again to bind to its configured address:port. ([#42272](https://github.com/kubernetes/kubernetes/pull/42272), [@marun](https://github.com/marun))

* The Prometheus metrics for the kube-apiserver for tracking incoming API requests and latencies...