- GCE Internal LoadBalancer sync loop will now release the ILB IP address upon sync failure. An error in ILB forwarding rule creation will no longer leak IP addresses. ([#97740](https://github.com/kubernetes/kubernetes/pull/97740), [@prameshj](https://github.com/prameshj)) [SIG Cloud Provider and Network]

- GCE Internal LoadBalancer sync loop will now release the ILB IP address upon sync failure. An error in ILB forwarding rule creation will no longer leak IP addresses. ([#97740](https://github.com/kubernetes/kubernetes/pull/97740), [@prameshj](https://github.com/prameshj)) [SIG Cloud Provider and Network]

## Dependencies

### Added
_Nothing has changed._

### Changed

- The Endpoints controller will now set the `endpoints.kubernetes.io/over-capacity` annotation to "warning" when an Endpoints resource contains more than 1000 addresses. In a future release, the controller will truncate Endpoints that exceed this limit. The EndpointSlice API can be used to support significantly larger number of addresses. ([#99975](https://github.com/kubernetes/kubernetes/pull/99975), [@robscott](https://github.com/robscott)) [SIG Apps and Network]

* [metrics-server addon] Restore connecting to nodes via IP addresses ([#76819](https://github.com/kubernetes/kubernetes/pull/76819), [@serathius](https://github.com/serathius))

- Added support to the kube-proxy nodePortAddresses / --nodeport-addresses option to
  accept the value "primary", meaning to only listen for NodePort connections
  on the node's primary IPv4 and/or IPv6 address (according to the Node object).
  This is strongly recommended, if you were not previously using
  --nodeport-addresses, to avoid surprising behavior.

- Fix a regression in kubenet  that  prevent pods to obtain ip addresses ([#87287](https://github.com/kubernetes/kubernetes/pull/87287), [@aojea](https://github.com/aojea)) [SIG Network and Node]

- Kube-proxy health check ports used to listen to `:<port>` for each of the services. This is not needed and opens ports in addresses the cluster user may not have intended. The PR limits listening to all node address which are controlled by `--nodeport-addresses` flag. if no addresses are provided then we default to existing behavior by listening to `:<port>` for each service ([#104742](https://github.com/kubernetes/kubernetes/pull/104742), [@khenida...

* Only use the first IP address got from instance metadata. This is because Azure CNI would set up a list of IP addresses in instance metadata, while only the first one is the Node's IP. ([#71736](https://github.com/kubernetes/kubernetes/pull/71736), [@feiskyer](https://github.com/feiskyer))

* Prevent kubelet from getting wedged if initialization of modules returns an error. ([#59020](https://github.com/kubernetes/kubernetes/pull/59020), [@brendandburns](https://github.com/brendandburns))

* kubelets that specify `--cloud-provider` now only report addresses in Node status as determined by the cloud provider ([#65594](https://github.com/kubernetes/kubernetes/pull/65594), [@liggitt](https://github.com/liggitt))
   * kubelet serving certificate rotation now reacts to changes in reported node addresses, and will request certificates for addresses set by an external cloud provider