- During device plugin allocation, resources requested by the pod can only be allocated if the device plugin has registered itself to kubelet AND healthy devices are present on the node to be allocated. If these conditions are not sattisfied, the pod would fail with `UnexpectedAdmissionError` error. ([#116337](https://github.com/kubernetes/kubernetes/pull/116337), [@swatisehgal](https://...

/davidz627))

* The LocalStorageCapacityIsolation feature is beta and enabled by default.  The LocalStorageCapacityIsolation feature added a new resource type ResourceEphemeralStorage "ephemeral-storage" so that this resource can be allocated, limited, and consumed as the same way as CPU/memory. All the features related to resource management (resource request/limit, quota, limitrange) are available for local ephemeral storage. This local ephemeral storage represents the storage for root...

- Replace scale up forbidden window with disregarding CPU samples collected when pod was initializing. ([#67252](https://github.com/kubernetes/kubernetes/pull/67252), [@jbartosik](https://github.com/jbartosik))

- When using `kubectl replace` (or the equivalent API call) on a Service, the caller no longer needs to do a read-modify-write cycle to fetch the allocated values for `.spec.clusterIP` and `.spec.ports[].nodePort`.  Instead the API server will automatically carry these forward from the original object when the new object does not specify them. ([#104673](https://github.com/kubernetes/kubernetes/pull/104673),...

- Added a "reason" field to the "image_garbage_collected_total" metric, allowing administrators to differentiate between images that were collected for reasons "age" vs "space".
   ([#123345](https://github.com/kubernetes/kubernetes/pull/123345), [@haircommander](https://github.com/haircommander))

- When using `kubectl replace` (or the equivalent API call) on a Service, the caller no longer needs to do a read-modify-write cycle to fetch the allocated values for `.spec.clusterIP` and `.spec.ports[].nodePort`.  Instead the API server will automatically carry these forward from the original object when the new object does not specify them. ([#104672](https://github.com/kubernetes/kubernetes/pull/104672),...

- When using `kubectl replace` (or the equivalent API call) on a Service, the caller no longer needs to do a read-modify-write cycle to fetch the allocated values for `.spec.clusterIP` and `.spec.ports[].nodePort`.  Instead the API server will automatically carry these forward from the original object when the new object does not specify them. ([#104674](https://github.com/kubernetes/kubernetes/pull/104674),...

 ([#42873](https://github.com/kubernetes/kubernetes/pull/42873),  [@xilabao](https://github.com/xilabao))
* You can now display requests/limits of extended resources in node allocated resources. ([#46079](https://github.com/kubernetes/kubernetes/pull/46079), [@xiangpengzhao](https://github.com/xiangpengzhao))

        `tf.function`-local variable: during each execution of the
        `tf.function`, the variable will be created and then disposed, similar
        to a local (i.e. stack-allocated) variable in C/C++. Currently
        `experimental_enable_variable_lifting=False` only works on non-XLA
        devices (e.g. under `@tf.function(jit_compile=False)`).

*   TF SavedModel:

* Support German cloud for azure disk mount feature ([#50673](https://github.com/kubernetes/kubernetes/pull/50673), [@clement-buchart](https://github.com/clement-buchart))
* BugFix: Exited containers are not Garbage Collected by the kubelet while the pod is running ([#53167](https://github.com/kubernetes/kubernetes/pull/53167), [@dashpole](https://github.com/dashpole))