- In clusters where the DryRun feature is enabled, dry-run requests will go through the normal admission chain. Because of this, ImagePolicyWebhook authors should especially make sure that their webhooks do not rely on side effects. ([#66391](https://github.com/kubernetes/kubernetes/pull/66391), [@jennybuckley](https://github.com/jennybuckley))

deleted pods may fail to terminate until the kubelet is restarted when the container runtime returns an error during termination. We have strengthened testing for runtime failures and now perform a more rigorous reconciliation to ensure static pods (especially those that use fixed UIDs) are restarted.  As a side effect of these changes static pods will be restarted with lower latency than before (2s vs 4s, on average) and rapid updates to pod configuration should take effect sooner.
  
  A...

  Previously, when unschedulable Pods are updated, the scheduler always put Pods back to activeQ/backoffQ. But, actually not all updates to Pods make Pods schedulable, especially considering many scheduling constraints nowadays are immutable.

- Disabled the `SchedulerAsyncAPICalls` feature gate to mitigate a bug where its interaction with asynchronous preemption could degrade `kube-scheduler` performance, especially under high `kube-apiserver` load. ([#134400](https://github.com/kubernetes/kubernetes/pull/134400), [@macsko](https://github.com/macsko))

### `tf.keras`:

*   Improvements from the Functional API refactoring:
    *   Functional model construction does not need to maintain a global
        workspace graph, removing memory leaks especially when building many
        models or very large models.
    *   Functional model construction should be ~8-10% faster on average.
    *   Functional models can now contain non-symbolic values in their call

  to fetch the OpenAPI may be slower. ([#118808](https://github.com/kubernetes/kubernetes/pull/118808), [@Jefftree](https://github.com/Jefftree))
- Shrank the `OpenAPI v2` spec by more than 50%, especially for less CPU resource consumption. ([#118204](https://github.com/kubernetes/kubernetes/pull/118204), [@sttts](https://github.com/sttts))

- Kubernetes will try to acquire the iptables lock every 100 msec during 5 seconds instead of every second. This is especially useful for environments using kube-proxy in iptables mode with a high churn rate of services. ([#85771](https://github.com/kubernetes/kubernetes/pull/85771), [@aojea](https://github.com/aojea)) [SIG Network]

* Kubeadm is currently omitting etcd certificates in a self-hosted deployment; this will be fixed in a point relelase. ([#61322](https://github.com/kubernetes/kubernetes/issues/61322))

* Some users, especially those with very large clusters, may see higher memory usage by the kube-controller-manager in 1.10. ([#61041](https://github.com/kubernetes/kubernetes/issues/61041))

  - This feature prevents the node and scheduler from oversubscribing resources
    that are still being used by a pod in the process of being cleaned up.
  - This feature can result in increase of Pod Deletion latency especially if the
    pod has a large filesystem footprint.

### RBAC
* **[beta]** RBAC API is promoted to v1beta1 (rbac.authorization.k8s.io/v1beta1), and defines default roles for control plane, node, and controller components.

- Eliminated work when creating Services or understanding port purposes, especially for external resources deployed via Helm charts. ([#133018](https://github.com/kubernetes/kubernetes/pull/133018), [@rushmash91](https://github.com/rushmash91))