- When registering with a 1.17+ API server, a CustomResourceDefinition conversion webhook can now request that only `v1` ConversionReview requests be sent to them. Previously, conversion webhooks were required to support receiving `v1beta1` ConversionReview requests as well for compatibility...

- Fixed a bug where the device resources allocated to an init container, with `containerRestartPolicy` of `Always`, were erroneously reused by a regular container. ([#120461](https://github.com/kubernetes/kubernetes/pull/120461),...

   ([#123273](https://github.com/kubernetes/kubernetes/pull/123273),...

- Fixed a suboptimal scheduler preemption behavior where potential preemption victims were violating Pod Disruption Budgets. ([#128307](https://github.com/kubernetes/kubernetes/pull/128307), [@NoicFank](https://github.com/NoicFank)) [SIG Scheduling]
- Fixed an issue in the kubelet that showed when writeable layers and read-only layers were at different paths within the same mount.

- Static pods were taking extra time to be restarted after being updated.  Static pods that are waiting to restart were not correctly counted in `kubelet_working_pods`. ([#116995](https://github.com/kubernetes/kubernetes/pull/116995), [@smarterclayton](https://github.com/smarterclayton)) [SIG Node]

- The fallback behavior of the Downward API's `resourceFieldRef` field was updated to account for pod-level resources: if container-level limits were not set, pod-level limits were now used before falling back to node allocatable resources. ([#132605](https://github.com/kubernetes/kubernetes/pull/132605), [@toVersus](https://github.com/toVersus)) [SIG Node, Scheduling and Testing]

* The following deprecated flags were removed from `kube-controller-manager`:

  * `replication-controller-lookup-cache-size`
  * `replicaset-lookup-cache-size`
  * `daemonset-lookup-cache-size`

  Don't use these flags. Using deprecated flags causes the server to print a warning. Using a removed flag causes the server to abort the startup.

* The following deprecated flags were removed from `kubelet`:

  
    - Traffic to a service IP now starts getting rejected (as opposed to
      merely dropped) as soon as there are no longer any *usable*
      endpoints, rather than waiting until all of the terminating
      endpoints have terminated even when those terminating endpoints
      were not being used.
  
    - Chains for endpoints that won't be used are no longer output to

  returns `nil`. (It now returns a pointer to a zero-valued object, like the real
  clients do.) This will break some downstream unit tests that were testing
  `result == nil` rather than `err != nil`, and in some cases may expose bugs

- e2e storage tests now run faster and are easier to read ([#72434](https://github.com/kubernetes/kubernetes/pull/72434), [@pohly](https://github.com/pohly))
- `e2e.test` now rejects unknown `--provider` values instead of merely warning about them. An empty provider name is not accepted anymore and was replaced by `skeleton` (a provider with no special behavior). ([#73402](https://github.com/kubernetes/kubernetes/pull/73402), [@pohly](https://github.com/pohly))