* The Prometheus metrics for the kube-apiserver for tracking incoming API requests and latencies now return the subresource label for correctly attributing the type of API call. ([#46354](https://github.com/kubernetes/kubernetes/pull/46354), [@smarterclayton](https://github.com/smarterclayton))

* Fix Service Update on LoadBalancerSourceRanges Field ([#37720](https://github.com/kubernetes/kubernetes/pull/37720), [@freehan](https://github.com/freehan))
* Bug fix. Incoming UDP packets not reach newly deployed services ([#32561](https://github.com/kubernetes/kubernetes/pull/32561), [@zreigz](https://github.com/zreigz))

### Other (Cleanup or Flake)

- Kube-apiserver: sets an upper-bound on the lifetime of idle keep-alive connections and time to read the headers of incoming requests ([#103958](https://github.com/kubernetes/kubernetes/pull/103958), [@liggitt](https://github.com/liggitt)) [SIG API Machinery and Node]

## Dependencies

### Added
_Nothing has changed._

### Changed

- `plugin_evaluation_total` metric supports prescore/score extension point.
  The metric doesn't get incremented when the prescore/score plugin has nothing to do with an incoming pod. ([#118025](https://github.com/kubernetes/kubernetes/pull/118025), [@AxeZhan](https://github.com/AxeZhan))

### Documentation

*   `tf.debugging`
    *   Add `tf.debugging.enable_check_numerics()` and
        `tf.debugging.disable_check_numerics()` to help debugging the root
        causes of issues involving infinities and `NaN`s.
*   `tf.distribute`
    *   Custom training loop support on TPUs and TPU pods is available through
        `strategy.experimental_distribute_dataset`,

* The Prometheus metrics for the kube-apiserver for tracking incoming API requests and latencies now return the `subresource` label for correctly attributing the type of API call. ([#46354](https://github.com/kubernetes/kubernetes/pull/46354), [@smarterclayton](https://github.com/smarterclayton))