- FIX: prevent apiserver from panicking when failing to load audit webhook config file ([#88879](https://github.com/kubernetes/kubernetes/pull/88879), [@JoshVanL](https://github.com/JoshVanL)) [SIG API Machinery and Auth]

  So she was considering in her own mind (as well as she could,
for the hot day made her feel very sleepy and stupid), whether
the pleasure of making a daisy-chain would be worth the trouble
of getting up and picking the daisies, when suddenly a White
Rabbit with pink eyes ran close by her.

  There was nothing so VERY remarkable in that; nor did Alice
think it so VERY much out of the way to hear the Rabbit say to

- DRA: when asking for "All" devices on a node, Kubernetes <= 1.32 proceeded to schedule pods onto nodes with no devices by not allocating any devices for those pods. Kubernetes 1.33 changes that to only picking nodes which have at least one device. Users who want the "proceed with scheduling also without devices" semantic can use the upcoming prioritized list feature with one sub-request for "all" devices and a second alternative with "count: 0". ([#129...

either a fixed-length computed hash, or parsed from a // field which has the same length limitation. // // Another reasonable approach might be to return a randomized slice if // we encounter an error, which would break the connection, but avoid // panicking. This would perhaps be safer but significantly more // confusing to users. panic("tls13: label or context too long") } hkdfLabel := make([]byte, 0, 2+1+len("tls13 ")+len(label)+1+len(context)) hkdfLabel = byteorder.BEAppendUint16(hkdfLabel, uint16(length))...

- When a CNI plugin returns dual-stack pod IPs, kubelet will now try to respect the
  "primary IP family" of the cluster by picking a primary pod IP of the same family
  as the (primary) node IP, rather than assuming that the CNI plugin returned the IPs
  in the order the administrator wanted (since some CNI plugins don't allow

nfiguration/assign-pod-node/). Node affinity/anti-affinity allow you to specify rules for restricting which node(s) a pod can schedule onto, based on the labels on the node. Pod affinity/anti-affinity allow you to specify rules for spreading and packing pods relative to one another, across arbitrary topologies (node, zone, etc.) These affinity rules are now be specified in a new-in-1.6 `affinity` field of the PodSpec. Kubernetes 1.6 continues to support the alpha `scheduler.alpha.kubernetes.io/affinity`...

        [example](https://github.com/tensorflow/models/blob/master/official/modeling/grad_utils.py)
        usage.
    *   Experimental support of
        [all reduce gradient packing](https://www.tensorflow.org/api_docs/python/tf/distribute/experimental/CollectiveHints)
        to allow overlapping gradient aggregation with backward path
        computation.