- Fix handling of aws-load-balancer-security-groups annotation. Security-Groups assigned with this annotation are no longer modified by kubernetes which is the expected behaviour of most users. Also no unnecessary Security-Groups are created anymore if this annotation is used. ([#88691](https://github.com/kubernetes/kubernetes/pull/88691), [@Elias481](https://github.com/Elias481)) [SIG Cloud Provider]

- Updated the CRI-O socket path, so users who configure kubelet to use a location like `/run/crio/crio.sock` don't see strange behaviour from CRI stats provider. ([#118704](https://github.com/kubernetes/kubernetes/pull/118704), [@dgl](https://github.com/dgl))

    `LSTMStateTuple`. The new behavior leads to proper dropout behavior for
    LSTMs and stacked LSTMs. This bug fix follows recommendations from published
    literature, but is a behavioral change. State dropout behavior may be
    customized via the new `dropout_state_filter_visitor` argument.
*   Removed `tf.contrib.training.python_input`. The same behavior, in a more

- Fix handling of aws-load-balancer-security-groups annotation. Security-Groups assigned with this annotation are no longer modified by kubernetes which is the expected behaviour of most users. Also no unnecessary Security-Groups are created anymore if this annotation is used. ([#88689](https://github.com/kubernetes/kubernetes/pull/88689), [@Elias481](https://github.com/Elias481)) [SIG Cloud Provider]

implementation is changed to not return an error when the specified target can be detected as not a mount point. On Linux, the behavior of detecting a mount point depends on `umount` command is validated when the mounter is created. Additionally, mount point checks will be skipped in CleanupMountPoint/CleanupMountWithForce if the mounter's Unmount having the changed behavior of not returning error when target is not a mount point. ([#109676](https://github.com/kubernetes/kubernetes/pull/109676), [...

* Default mount propagation has changed from `HostToContainer` ("rslave" in Linux terminology), as it was in 1.10, to `None` ("private") to match the behavior in 1.9 and earlier releases; `HostToContainer` as a default caused regressions in some pods. If you are relying on this behavior you will need to set it explicitly. ([#62462](https://github.com/kubernetes/kubernetes/pull/62462), [@jsafrane](https://github.com/jsafrane))

    - [CSI Health Monitoring](#csi-health-monitoring)
  - [Known Issues](#known-issues)
    - [<code>TopologyAwareHints</code> feature falls back to default behavior](#topologyawarehints-feature-falls-back-to-default-behavior)
  - [Urgent Upgrade Notes](#urgent-upgrade-notes)
    - [(No, really, you MUST read this before you upgrade)](#no-really-you-must-read-this-before-you-upgrade)

  * Marking a StorageClass as default makes it so that even a PersistentVolumeClaim without a StorageClass specified will trigger dynamic provisioning (instead of binding to an existing pool of PVs).

  This is strongly recommended, if you were not previously using
  --nodeport-addresses, to avoid surprising behavior.
  (This behavior is enabled by default with the nftables backend; you would
  need to explicitly request `--nodeport-addresses 0.0.0.0/0,::/0` there to get

  To request this behaviour, your client software must enable the `WatchListClient` client-go feature gate. Additionally, streaming is only available if supported by the cluster; the API server that you connect to must also support streaming.