- The kubelet, when using --cloud-provider=external can use the --node-ip flag with one of the unspecified addresses 0.0.0.0 or ::, to create the Node with the IP of the default gateway of the corresponding IP family and then delegating the responsibility to the external cloud provider. This solve the bootstrap problems of out of tree cloud providers that are deployed as...

- Expand scheduler priority functions and scheduling framework plugins' node score range to [0, 100]. Note: this change is internal and does not affect extender and RequestedToCapacityRatio custom priority, which are still expected to provide a [0, 10] range. ([#83522](https://github.com/kubernetes/kubernetes/pull/83522), [@draveness](https://g...

parameter in kube-proxy is misleading, no port is opened with this address. Instead it is translated internally to "nodeIP". The nodeIPs for both families are now taken from the Node object if `--bind-address` is unspecified or set to the "any" address (0.0.0.0 or ::). It is recommended to leave `--bind-address` unspecified, and in particular avoid to set it to localhost (127.0.0.1 or ::1) ([#119525](https://github.com/kubernetes/kubernetes/pull/119525), [@uablrek](https://github.com/uablrek)) [SIG Network...

				fmt.Sprintf("bytes=-%d", objLen/2),
				// Read middle half of object
				fmt.Sprintf("bytes=%d-%d", objLen/4, objLen*3/4),
				// Read 100MiB of the object from the beginning
				fmt.Sprintf("bytes=%d-%d", 0, 100*humanize.MiByte),
				// Read 100MiB of the object from the end
				fmt.Sprintf("bytes=-%d", 100*humanize.MiByte),
			}
			for _, rangeHdr := range rangeHdrs {
				mkGetReq(oi, rangeHdr, caseNumber, sf)

					// Log if we took a lot of time.
					logger.Info("Site replication healing refresh took %.2fs", took)
				}

				// wait for 200 millisecond, if we are experience lot of I/O
				waitForLowIO(runtime.GOMAXPROCS(0), 200*time.Millisecond, currentHTTPIO)
			}
			healTimer.Reset(siteHealTimeInterval)

		case <-ctx.Done():
			return
		}
	}
}

type srBucketStatsSummary struct {

- Add --permit-address-sharing flag to `kube-apiserver` to listen with `SO_REUSEADDR`. While allowing to listen on wildcard IPs like 0.0.0.0 and specific IPs in parallel, it avoids waiting for the kernel to release socket in `TIME_WAIT` state, and hence, considerably reducing `kube-apiserver` restart times under certain conditions. ([#93861](https://github.com/kubernetes/kubernetes/pull/93861),...

- Apiserver will now reject connection attempts to `0.0.0.0/::` when handling a proxy subresource request. ([#107402](https://github.com/kubernetes/kubernetes/pull/107402), [@anguslees](https://github.com/anguslees))

* kubeadm: Create /var/lib/etcd with correct permissions (0700) by default. ([#71885](https://github.com/kubernetes/kubernetes/pull/71885), [@dims](https://github.com/dims))

  * Tensorflow will continue to support NumPy 1.26 until 2025, aligning with community standard deprecation timeline [here](https://scientific-python.org/specs/spec-0000/).

* Hermetic CUDA support is added.

* Add RequestedToCapacityRatioPriority priority function. Function is parametrized with set of points mapping node utilization (0-100) to score (0-10). ([#63929](https://github.com/kubernetes/kubernetes/pull/63929), [@losipiuk](https://github.com/losipiuk))