- Terminated Pods on a node will not be re-admitted on kubelet restart. This fixes the problem of Completed Pods awaiting for the finalizer marked as Failed after the kubelet restart. ([#127210](https://github.com/kubernetes/kubernetes/pull/127210), [@SergeyKanzhelev](https://github.com/SergeyKanzhelev)) [SIG Node and Testing]

		// considered as Delete marker true to avoid listing such objects by
		// regular ListObjects() calls. However for delete replication this
		// ends up being a problem because "upon" a successful delete this
		// ends up creating a new delete marker that is spurious and unnecessary.
		//
		// Regression introduced by #14555 was reintroduced in #15564
		if versionFound {

    * - Update go version to 1.9.3.
* Backport Cloud CIDR allocator fixes to 1.9 ([#61797](https://github.com/kubernetes/kubernetes/pull/61797), [@satyasm](https://github.com/satyasm))
* openstack cinder detach problem is fixed if nova is shutdowned ([#56846](https://github.com/kubernetes/kubernetes/pull/56846), [@zetaab](https://github.com/zetaab))

- Revised the logic for DaemonSet rolling update to exclude nodes if scheduling constraints are not met.
  This eliminates the problem of rolling updates to a DaemonSet getting stuck around tolerations. ([#120792](https://github.com/kubernetes/kubernetes/pull/120792), [@mochizuki875](https://github.com/mochizuki875)) [SIG Apps and Testing]

		// call will report top level prefixes in deleted state, whereas spark/hadoop interpret this as non-empty
		// and throw a 404 exception. This is especially a problem for spark jobs overwriting the same partition
		// repeatedly. This workaround recursively lists the top 3 entries including delete markers to reflect the
		// correct state of the directory in the list results.

back and look at this code closely (wtf?)\n  // If the content of the popper grows once it's been positioned, it\n  // may happen that the popper gets misplaced because of the new content\n  // overflowing its reference element\n  // To avoid this problem, we provide two options (x and y), which allow\n  // the consumer to define the offset origin.\n  // If we position a popper on top of a reference element, we can set\n  // `x` to `top` to make the popper grow towards its top instead of\n  // its...

- Azure: check the nil error before invoke err.Error() ([#96623](https://github.com/kubernetes/kubernetes/pull/96623), [@nilo19](https://github.com/nilo19)) [SIG Cloud Provider]
- Bump node-problem-detector version to v0.8.5 to fix OOM detection in with Linux kernels 5.1+ ([#96716](https://github.com/kubernetes/kubernetes/pull/96716), [@tosi3k](https://github.com/tosi3k)) [SIG Cloud Provider, Scalability and Testing]

* Fixed parsing of fsType in AWS StorageClass parameters ([#75944](https://github.com/kubernetes/kubernetes/pull/75944), [@jsafrane](https://github.com/jsafrane))
* Node-Problem-Detector configuration is now decoupled from the Kubernetes release on GKE/GCE. ([#73288](https://github.com/kubernetes/kubernetes/pull/73288), [@wangzhen127](https://github.com/wangzhen127))

## Breaking Changes

*   `tf.keras`:
    *   In `tf.keras.applications` the name of the "top" layer has been
        standardized to "predictions". This is only a problem if your code
        relies on the exact name of the layer.
    *   Huber loss function has been updated to be consistent with other Keras
        losses. It now computes mean over the last axis of per-sample losses

- Bump node-problem-detector version to v0.8.5 to fix OOM detection in with Linux kernels 5.1+ ([#96716](https://github.com/kubernetes/kubernetes/pull/96716), [@tosi3k](https://github.com/tosi3k)) [SIG Cloud Provider, Scalability and Testing]