- [KCCM - service controller]: enabled connection draining for terminating pods upon node downscale by the cluster autoscaler. This is done by not reacting to the taint used by the cluster autoscaler to indicate that the node is going away soon, thus keeping the node referenced by the load balancer until the VM has been completely deleted. ([#115204](https://github.com/kubernetes/kubernetes/pull/115204), [@alexanderConstantinescu](https://github.com/alexanderConstantinescu))

* Dual stack services (Phase II of IPv6DualStack feature) are enabled via the IPVS proxier. iptables proxier does not support dualstack yet. Dualstack iptables proxier is WIP and should catchup soon. ([#82091](https://github.com/kubernetes/kubernetes/pull/82091), [@khenidak](https://github.com/khenidak))
    * to enable, kube-proxy must be have the following flags:
    * --proxy-mode=ipvs
    * --cluster-cidrs=<cidr>,<cidr>

  - in cases where the client specifies a timeout in the request URL, the overall request deadline is shortened now since the deadline is setup as soon as the request is received by the apiserver. ([#96901](https://github.com/kubernetes/kubernetes/pull/96901), [@tkashem](https://github.com/tkashem)) [SIG API Machinery and Testing]

  * Caveat: The alpha version no longer supports, and the beta version does not support, the "empty `podAffinityTerm.namespaces` list means all namespaces" behavior. In both alpha and beta it now means "same namespace as the pod...

  as a failure until that Pod is terminal (its `.status.phase` is `Failed` or `Succeeded`).
  
  However, the Job controller creates a replacement Pod as soon as the termination becomes apparent.
  Once the pod terminates, the Job controller evaluates `.backoffLimit` and `.podFailurePolicy` for the
  relevant Job, taking this now-terminated Pod into consideration.
  

      (rather than continuing until the endpoint is fully deleted).
  
    - Traffic to a service IP now starts getting rejected (as opposed to
      merely dropped) as soon as there are no longer any *usable*
      endpoints, rather than waiting until all of the terminating
      endpoints have terminated even when those terminating endpoints
      were not being used.