- Dynamic resource allocation: when a claim uses "wait for first consumer" allocation (the default), then it will now get deallocated after it was used by a pod. That ensures that the next pod isn't affected by previous scheduling decision and that resources are not kept allocated unless really needed. If keeping a claim allocated is desired, use "immediate allocation." ([#118936](https://github.com/kubernetes/kubernetes/pull/118936), [@pohl...

   * #aboutToAcquire(CycleDetectingLock)} and {@link #lockStateChanged(CycleDetectingLock)}.
   */
  // This is logically a Set, but an ArrayList is used to minimize the amount
  // of allocation done on lock()/unlock().
  private static final ThreadLocal<List<LockGraphNode>> acquiredLocks =
      new ThreadLocal<List<LockGraphNode>>() {
        @Override
        protected List<LockGraphNode> initialValue() {

- DRA kubelet: gRPC API graduated to v1, v1beta1 was deprecated starting in 1.34. Updating DRA drivers to the `k8s.io/dynamic-resource-allocation/kubeletplugin` helper from 1.34 added support for both API versions. ([#132700](https://github.com/kubernetes/kubernetes/pull/132700), [@pohly](https://github.com/pohly)) [SIG Node and Testing]

    int size = thisList.size();
    if (size != otherList.size()) {
      return false;
    }
    if (thisList instanceof RandomAccess && otherList instanceof RandomAccess) {
      // avoid allocation and use the faster loop
      for (int i = 0; i < size; i++) {
        if (!Objects.equals(thisList.get(i), otherList.get(i))) {
          return false;
        }
      }
      return true;

- Dynamic Resource Allocation (DRA): client-side validation of a ResourceHandle would have accepted a missing DriverName, whereas server-side validation then would have raised an error. ([#124075](https://github.com/kubernetes/kubernetes/pull/124075), [@pohly](https://github.com/pohly))

		bytebufferpool.Put(sb)
	}()

	return pathJoinBuf(sb, elem...)
}

// pathJoinBuf - like path.Join() but retains trailing SlashSeparator of the last element.
// Provide a string builder to reduce allocation.
func pathJoinBuf(dst *bytebufferpool.ByteBuffer, elem ...string) string {
	trailingSlash := len(elem) > 0 && hasSuffixByte(elem[len(elem)-1], SlashSeparatorChar)
	dst.Reset()
	added := 0
	for _, e := range elem {

   ([#123508](https://github.com/kubernetes/kubernetes/pull/123508), [@saschagrunert](https://github.com/saschagrunert))
- Dynamic Resource Allocation: DRA drivers can now use "structured parameters" to let the scheduler handle claim allocation.
   ([#123516](https://github.com/kubernetes/kubernetes/pull/123516), [@pohly](https://github.com/pohly))

        }
      }
      return new EntrySetImpl();
    }

    @Override
    public void forEach(BiConsumer<? super K, ? super V> action) {
      checkNotNull(action);
      // avoids allocation of entries
      backingSet().forEach(k -> action.accept(k, function.apply(k)));
    }
  }

  static <K extends @Nullable Object, V extends @Nullable Object>

        previous gradient implementation.
    *   Added a new session config setting `internal_fragmentation_fraction`,
        which controls when the BFC Allocator needs to split an oversized chunk
        to satisfy an allocation request.
    *   Added `tf.get_current_name_scope()` which returns the current full name
        scope string that will be prepended to op names.
*   `tf.data`:

          // We might, somehow, be able to reorder values in-place.  But it doesn't seem like
          // there's a way around creating the separate sortedKeys array, and if we're allocating
          // one array of size n, we might as well allocate two -- to say nothing of the allocation
          // done in Arrays.sort.
          for (int i = 0; i < size; i++) {
            // We're careful to put only K instances in.