- Fixed an error processing watch events when running skewed apiservers ([#73482](https://github.com/kubernetes/kubernetes/pull/73482), [@liggitt](https://github.com/liggitt))
- jsonpath expressions containing `[start:end:step]` slice are now evaluated correctly ([#73149](https://github.com/kubernetes/kubernetes/pull/73149), [@liggitt](https://github.com/liggitt))

### Bug or Regression

- Fixed a regression since 1.27.0 in scheduler framework when running score plugins. 
  The `skippedScorePlugins` number might be greater than `enabledScorePlugins`, 
  so when initializing a slice the cap(len(skippedScorePlugins) - len(enabledScorePlugins)) is negative, 
  which is not allowed. ([#121666](https://github.com/kubernetes/kubernetes/pull/121666), [@kerthcet](https://github.com/kerthcet)) [SIG Scheduling]

- Fixed a bug that mistake use newObj as oldObj in endpoint slice update. ([#92339](https://github.com/kubernetes/kubernetes/pull/92339), [@fatkun](https://github.com/fatkun)) [SIG Apps and Network]

- Endpoint slices mirroring controller no longer mirrors the last-applied-configuration annotation created by kubectl to updated endpoint slices ([#102731](https://github.com/kubernetes/kubernetes/pull/102731), [@sharmarajdaksh](https://github.com/sharmarajdaksh)) [SIG API Machinery, Apps, Cloud Provider, Network, Release, Scheduling,...

- Azure: avoid setting cached Sku when updating VMSS and VMSS instances ([#102005](https://github.com/kubernetes/kubernetes/pull/102005), [@feiskyer](https://github.com/feiskyer)) [SIG Cloud Provider]
- Fix a bug on the endpoint slices mirroring controller where endpoint NotReadyAddresses were mirrored as Ready to the corresponding EndpointSlice ([#102683](https://github.com/kubernetes/kubernetes/pull/102683), [@aojea](https://github.com/aojea)) [SIG Apps and Network]

- Significant kube-proxy performance improvements when using Endpoint Slices at scale. ([#83206](https://github.com/kubernetes/kubernetes/pull/83206), [@robscott](https://github.com/robscott))

### Node

- DRA drivers: the resource slice controller sometimes didn't react properly when kubelet or someone else deleted a recently created ResourceSlice. It incorrectly assumed that the ResourceSlice still exists and didn't recreate it. ([#132683](https://github.com/kubernetes/kubernetes/pull/132683),...

        `tf.sparse.cross`/`tf.ragged.cross` directly.
    *   Added additional `standardize` and `split` modes to `TextVectorization`:
        *   `standardize="lower"` will lowercase inputs.
        *   `standardize="string_punctuation"` will remove all punctuation.
        *   `split="character"` will split on every unicode character.
    *   Added an `output_mode` argument to the `Discretization` and `Hashing`

 * kubeadm: The Token-related fields in the `MasterConfiguration` object have now been refactored. Instead of the top-level `.Token`, `.TokenTTL`, `.TokenUsages`, `.TokenGroups` fields, there is now a `BootstrapTokens` slice of `BootstrapToken` objects that support the same features under the `.Token`, `.TTL`, `.Usages`, `.Groups` fields. ([#64408](https://github.com/kubernetes/kubernetes/pull/64408), [@luxas](https://github.com/luxas))

- Fixed a regression since `v1.27.0` in the scheduler framework when running score plugins. 
  The `skippedScorePlugins` number might be greater than `enabledScorePlugins`, 
  so when initializing a slice the `cap(len(skippedScorePlugins) - len(enabledScorePlugins))` is negative, 
  which is not allowed. ([#121632](https://github.com/kubernetes/kubernetes/pull/121632), [@kerthcet](https://github.com/kerthcet))