* when updating a service or when node changes occur.
* Allow for watching objects larger than 1MB given etcd accepts objects of size up to 1.5MB ([#72053](https://github.com/kubernetes/kubernetes/pull/72053), [@wojtek-t](https://github.com/wojtek-t))

- Fixes an issue that can result in namespaced custom resources being orphaned when their namespace is deleted, if the CRD defining the custom resource is removed concurrently with namespaces being deleted, then recreated. ([#93790](https://github.com/kubernetes/kubernetes/pull/93790), [@liggitt](https://github.com/liggitt)) [SIG API Machinery and Apps]

- Fix: remove VMSS and VMSS instances from SLB backend pool only when necessary ([#105839](https://github.com/kubernetes/kubernetes/pull/105839), [@nilo19](https://github.com/nilo19))
- Fix: skip `instance not found` when decoupling VMSSs from LB ([#105666](https://github.com/kubernetes/kubernetes/pull/105666), [@nilo19](https://github.com/nilo19))

  - Services that have been set to IPFamilyPolicy = PreferDualstack will be upgraded when the service object is updated. e.g., when a user change a label.
  
  This behavior will change to:

- Warn instead of fail when creating Roles and ClusterRoles with custom verbs via kubectl ([#92492](https://github.com/kubernetes/kubernetes/pull/92492), [@eddiezane](https://github.com/eddiezane)) [SIG CLI]

- Sometimes, the scheduler incorrectly placed a pod in the "unschedulable" queue instead of the "backoff" queue. This happened when some plugin previously declared the pod as "unschedulable" and then in a later attempt encounters some other error. Scheduling of that pod then got delayed by up to five minutes, after which periodic flushing moved the pod back into the "active" queue. ([#120334](https://github.com/kubernetes/kubernetes/pull/120334), ...

systems, a mounter's Unmount 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/k...

* Fix panic in ControllerManager on GCE when it has a problem with creating external loadbalancer healthcheck ([#52646](https://github.com/kubernetes/kubernetes/pull/52646), [@gmarek](https://github.com/gmarek))

* Retry 'connection refused' errors when setting up clusters on GCE. ([#57394](https://github.com/kubernetes/kubernetes/pull/57394), [@mborsz](https://github.com/mborsz))
* Retry 'connection refused' errors when setting up clusters on GCE. ([#57394](https://github.com/kubernetes/kubernetes/pull/57394), [@mborsz](https://github.com/mborsz))

* Limit use of tags when calling EC2 API to prevent API throttling for very large clusters ([#76749](https://github.com/kubernetes/kubernetes/pull/76749), [@mcrute](https://github.com/mcrute))