- When `SchedulerQueueingHint` is enabled,
  the scheduler's in-tree plugins now subscribe to specific node events to decide whether to requeue Pods.
  This allows the scheduler to handle cluster events faster with less memory.

  Specific node events include updates to taints, tolerations or allocatable.

    *   Unary GPU kernels: Abs, Atanh, Acos, Acosh, Asin, Asinh, Atan, Cos, Cosh, Sin, Sinh, Tan, Tanh.
    *   Binary GPU kernels: AddV2, Sub, Div, DivNoNan, Mul, MulNoNan, FloorDiv, Equal, NotEqual, Greater, GreaterEqual, LessEqual, Less.

* `tf.lite`
    * Add experimental supports conversion of models that may be larger than 2GB before buffer deduplication

### Bug Fixes and Other Changes

- Count pod overhead against an entity's ResourceQuota ([#99600](https://github.com/kubernetes/kubernetes/pull/99600), [@gjkim42](https://github.com/gjkim42)) [SIG API Machinery and Node]
- EndpointSlice controller is now less likely to emit FailedToUpdateEndpointSlices events. ([#100146](https://github.com/kubernetes/kubernetes/pull/100146), [@robscott](https://github.com/robscott)) [SIG Apps and Network]

* removed unsafe double RLock in cpumanager ([#62464](https://github.com/kubernetes/kubernetes/pull/62464), [@choury](https://github.com/choury))
* Fix in vSphere Cloud Provider to handle upgrades from kubernetes version less than v1.9.4 to v1.9.4 and above. ([#62919](https://github.com/kubernetes/kubernetes/pull/62919), [@abrarshivani](https://github.com/abrarshivani))

The StorageProtection feature, which prevents deletion of PVCs while Pods are still using them and of PVs while still bound to a PVC, is now generally available, and volume resizing, which lets you increase size of a volume after a Pod restarts is now beta, which means it is on by default.

New alpha features include:

     * the expression begins with one or more '?'s then exactly that
     * many characters will be matched whereas if it ends with '?'s
     * it will match that many characters <i>or less</i>.
     * <p>
     * Wildcard expressions will not filter workgroup names or server names.
     *
     * <blockquote>
     *
     * <pre>
     * winnt&gt; ls c?o*

		// we multiply by the number of sets in the pool.
		// This will compensate for differences in set sizes
		// when choosing destination pool.
		// Different set sizes are already compensated by less disks.
		available *= uint64(nSets[i])

		serverPools[i] = poolAvailableSpace{
			Index:      i,
			Available:  available,
			MaxUsedPct: maxUsedPct,
		}
	}
	return serverPools
}

*[gcmBlockSize]byte, nonce []byte, productTable *[256]byte) { if len(nonce) == gcmStandardNonceSize { copy(counter[:], nonce) counter[gcmBlockSize-1] = 1 } else { var hash [16]byte paddedGHASH(&hash, nonce, productTable) lens := gcmLengths(0, uint64(len(nonce))*8) paddedGHASH(&hash, lens[:], productTable) copy(counter[:], hash[:]) } } // counterCrypt encrypts in using AES in counter mode and places the result // into out. counter is the initial count value and will be updated with the next // count value....

     * the expression begins with one or more '?'s then exactly that
     * many characters will be matched whereas if it ends with '?'s
     * it will match that many characters <i>or less</i>.
     * <p>
     * Wildcard expressions will not filter workgroup names or server names.
     *
     * <blockquote><pre>
     * winnt> ls c?o*

	// Parse the bucket modtime
	creationTime, err := time.Parse(iso8601TimeFormat, createdBucket.CreationDate)
	c.Assert(err, nil)

	// Check if bucket modtime is consistent (not less than current time and not late more than 5 minutes)
	timeNow := time.Now().UTC()
	c.Assert(creationTime.Before(timeNow), true)
	c.Assert(timeNow.Sub(creationTime) < time.Minute*5, true)
}