- The disk-cached discovery client is moved from k8s.io/client-go/discovery to k8s.io/client-go/discovery/cached/disk.
The memory-cached discovery client is moved from k8s.io/client-go/discovery/cached to k8s.io/client-go/discovery/cached/memory.
([#72214](https://github.com/kubernetes/kubernetes/pull/72214), [@caesarxuchao](https://github.com/caesarxuchao))

filtering on fields was incorrect. If watching objects with a filter, when an update was made that no longer matched the filter a DELETE event was correctly sent. However, the object that was returned by that delete was not the (correct) version before the update, but instead, the newer version. That meant the new object was not matched by the filter. This was a regression from behavior between cached watches on the server side and uncached watches, and thus broke downstream API clients. ([#46223](http...

- Clear pod binding cache on bind error to make sure stale pod binding cache will not be used. ([#71212](https://github.com/kubernetes/kubernetes/pull/71212), [@cofyc](https://github.com/cofyc))
- Fixed a scheduler panic due to internal cache inconsistency ([#71063](https://github.com/kubernetes/kubernetes/pull/71063), [@Huang-Wei](https://github.com/Huang-Wei))

- Extended resources backed by DRA feature allowed cluster operator to specify `extendedResourceName` in `DeviceClass`, and application operator to continue using extended resources in pod's requests to request for DRA devices matching the DeviceClass.
  

- Fixed a bug in the garbage collector controller which could block indefinitely due to a cache sync failure. This fix allows the garbage collector to eventually continue garbage collecting other resources if a given resource cannot be listed or watched. Any objects in the unsynced resource type with owner references with `blockOwnerDeletion: true` will not be known to the garbage collector. Use of `blockOwnerDeletion`...

- Bumped `ProcMountType` feature to on by default beta ([#130798](https://github.com/kubernetes/kubernetes/pull/130798), [@haircommander](https://github.com/haircommander)) [SIG Node]
- Calculated pod resources are now cached when adding pods to NodeInfo in the scheduler framework, improving performance when processing unschedulable pods. ([#129635](https://github.com/kubernetes/kubernetes/pull/129635), [@macsko](https://github.com/macsko)) [SIG Scheduling]

- Fix the bug PIP's DNS is deleted if no DNS label service annotation isn't set. ([#87310](https://github.com/kubernetes/kubernetes/pull/87310), [@nilo19](https://github.com/nilo19)) [SIG Cloud Provider]
- Fix: set nil cache entry based on old cache ([#87593](https://github.com/kubernetes/kubernetes/pull/87593), [@aramase](https://github.com/aramase)) [SIG Cloud Provider]

### SIG Scheduling

- The equivalence class cache has been redesigned to be a two level cache, resulting in a significant increase in scheduling throughput and performance. ([#65714](https://github.com/kubernetes/kubernetes/pull/65714), [@resouer](https://github.com/resouer))

* kubelet now accepts a --cni-cache-dir option, which defaults to /var/lib/cni/cache, where CNI stores cache files. ([#78908](https://github.com/kubernetes/kubernetes/pull/78908), [@dcbw](https://github.com/dcbw))

- Fix availability set cache in vmss cache ([#100110](https://github.com/kubernetes/kubernetes/pull/100110), [@CecileRobertMichon](https://github.com/CecileRobertMichon)) [SIG Cloud Provider]