* Fix issue where on re-registration of device plugin, `allocatable` was not getting updated. This issue makes devices invisible to the Kubelet if device plugin restarts. Only work-around, if this fix is not there, is to restart the kubelet and then start device plugin. ([#63118](https://github.com/kubernetes/kubernetes/pull/63118), [@vikaschoudhary16](https://github.com/vikaschoudhary16))

        decisions.
*   PluggableDevice
    *   Third-party devices can now connect to TensorFlow as plug-ins through
        [StreamExecutor C API](https://github.com/tensorflow/community/blob/master/rfcs/20200612-stream-executor-c-api.md).
        and
        [PluggableDevice](https://github.com/tensorflow/community/blob/master/rfcs/20200624-pluggable-device-for-tensorflow.md)
        interface.

* Fix issue where on re-registration of device plugin, `allocatable` was not getting updated. This issue makes devices invisible to the Kubelet if device plugin restarts. Only work-around, if this fix is not there, is to restart the kubelet and then start device plugin. ([#63118](https://github.com/kubernetes/kubernetes/pull/63118), [@vikaschoudhary16](https://github.com/vikaschoudhary16))

- Services implement a field selector for the ClusterIP and Type fields.
  Kubelet uses the fieldselector on Services to avoid watching for Headless Services and reduce the memory consumption. ([#123905](https://github.com/kubernetes/kubernetes/pull/123905), [@aojea](https://github.com/aojea)) [SIG Apps, Node and Testing]

- Fixed cleanup of raw block devices after kubelet restart. ([#83451](https://github.com/kubernetes/kubernetes/pull/83451), [@jsafrane](https://github.com/jsafrane))
- Add data cache flushing during unmount device for GCE-PD driver in Windows Server. ([#83591](https://github.com/kubernetes/kubernetes/pull/83591), [@jingxu97](https://github.com/jingxu97))

  * **alpha**: NVIDIA GPU support ([#24836](https://github.com/kubernetes/kubernetes/pull/24836), [@therc](https://github.com/therc))
  * Adding loadBalancer services and nodeports services to quota system

## Known Issues and Important Steps before Upgrading

/nolancon/website/blob/f4200307260ea3234540ef13ed80de325e1a7267/content/en/docs/tasks/administer-cluster/topology-manager.md) enables NUMA alignment of CPU and devices (such as SR-IOV VFs) that will allow your workload to run in an environment optimized for low-latency. Prior to the introduction of the Topology Manager, the CPU and Device Manager would make resource allocation decisions independent of each other. This could result in undesirable allocations on multi-socket systems, causing degraded...

- 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 sattsfied, the pod would fail with `UnexpectedAdmissionError` error. ([#117719](https://github.com/kuber...

  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
  TF_DeleteGraph(graph);
  TF_DeleteStatus(s);
}

// If `device` is non-empty, run Min op on that device.
// Otherwise run it on the default device (CPU).
void RunMinTest(const std::string& device, bool use_XLA) {
  TF_Status* s = TF_NewStatus();
  TF_Graph* graph = TF_NewGraph();

  // Make a placeholder operation.

## Changelog since v1.13.5

### Other notable changes

* Connections from Pods to Services with 0 endpoints will now ICMP reject immediately, rather than blackhole and timeout. ([#72534](https://github.com/kubernetes/kubernetes/pull/72534), [@thockin](https://github.com/thockin))