// causing the thread to be suspended for 3.5 seconds, and then park itself for 2 seconds after
    // being resumed. To avoid a flake in this scenario, calculate how long that thread actually
    // waited and assert based on that time. Empirically, the race where the thread ends up waiting
    // for 5.5 seconds happens about 2% of the time.

- kubeadm: Fixed a possible scenario where kubeadm can pull much newer control-plane images ([#69301](https://github.com/kubernetes/kubernetes/pull/69301), [@neolit123](https://github.com/neolit123))

- Do not remove the "batch.kubernetes.io/job-tracking" finalizer from a Pod, in a corner
  case scenario, when the Pod is controlled by an API object which is not a batch Job

- DRA: scheduling pods is up to 16x faster, depending on the scenario. Scheduling throughput depends a lot on cluster utilization. It is higher for lightly loaded clusters with free resources and gets lower when the cluster utilization increases. ([#127277](https://github.com/kubernetes/kubernetes/pull/127277), [@...

   ([#122022](https://github.com/kubernetes/kubernetes/pull/122022), [@sanposhiho](https://github.com/sanposhiho))
- QueueingHint implementation for `NodeAffinity` was reverted because potential scenarios were found where events that make Pods schedulable could be missed.

- Kubelet no longer watches `ConfigMaps` and `Secrets` for terminated pods, in worst scenario causing it to not be able to send other requests to kube-apiserver ([#74809](https://github.com/kubernetes/kubernetes/pull/74809), [@oxddr](https://github.com/oxddr))

- The scheduling queue didn't notice any extenders' failures, it could miss some cluster events,
  and it could end up Pods rejected by Extenders stuck in unschedulable pod pool in 5min in the worst-case scenario.

- QueueingHint implementation for NodeAffinity is reverted because we found potential scenarios where events that make Pods schedulable could be missed. ([#122327](https://github.com/kubernetes/kubernetes/pull/122327), [@sanposhiho](https://github.com/sanposhiho)) [SIG Scheduling]

 * </p>
 * <pre>
 * Map&lt;K, V&gt; syncMap = Collections.synchronizedMap(new LruHashMap&lt;&gt;(100));
 * </pre>
 * <p>
 * Alternatively, for high-concurrency scenarios, consider implementing a custom
 * LRU cache using {@link java.util.concurrent.ConcurrentHashMap} with an eviction strategy.
 * </p>
 *
 * @author koichik
 * @param <K> the key type
 * @param <V> the value type
 */

- `kubectl wait` ensures that observedGeneration >= generation to prevent stale state reporting. An example scenario can be found on CRD updates. ([#97408](https://github.com/kubernetes/kubernetes/pull/97408), [@KnicKnic](https://github.com/KnicKnic))

### Documentation