* Requests with the query parameter `?watch=` are treated by the API server as a request to watch, but authorization and metrics were not correctly identifying those as watch requests, instead grouping them as list calls. ([#48583](https://github.com/kubernetes/kubernetes/pull/48583), [@smarterclayton](https://github.com/smarterclayton))

- Fixed a race condition in `kube-proxy` when using LocalModeNodeCIDR, to avoid dropping Services traffic if the object node is recreated when `kube-proxy` is starting. ([#118499](https://github.com/kubernetes/kubernetes/pull/118499), [@aojea](https://github.com/aojea))

- The metrics `etcd_request_duration_seconds` and `etcd_bookmark_counts` now differentiate by group resource instead of object type, allowing unique entries per `CustomResourceDefinition`, instead of grouping them all under `*unstructured.Unstructured`. ([#112042](https://github.com/kubernetes/kubernetes/pull/112042), [@ncdc](https://github.com/ncdc))

- Improved memory usage of `kube-scheduler` by dropping the `.metadata.managedFields` field that `kube-scheduler` doesn't require. ([#119556](https://github.com/kubernetes/kubernetes/pull/119556), [@linxiulei](https://github.com/linxiulei))