- Fix: Bug in `kube-proxy` latency metrics to calculate only the latency value for the `Endpoints` that are created after it starts running. This is needed because all the `Endpoints` objects are processed on restarts, independently when they were. ([#100861](https://github.com/kubernetes/kubernetes/pull/100861), [@aojea](https://github.com/aojea))

- The RuntimeClass scheduler can now simplify steering Linux or Windows pods to appropriate nodes
- All Windows nodes now get the new label `node.kubernetes.io/windows-build` that reflects the Windows major, minor, and build number that are needed to match compatibility between Windows containers and Windows worker nodes.


## Deprecations and Removals

- When invoked with `-list-images`, the `e2e.test` binary now also lists the images that might be needed for storage tests. ([#108458](https://github.com/kubernetes/kubernetes/pull/108458), [@pohly](https://github.com/pohly))

                # By placing the middleware and the AsyncExitStack here, inside all
                # user middlewares, the same context is used.
                # This is currently not needed, only for closing files, but used to be
                # important when dependencies with yield were closed here.
                Middleware(AsyncExitStackMiddleware),
            ]
        )

- HTTP/2 connection health check is enabled by default in all Kubernetes clients to fix persistently broken connections (https://github.com/kubernetes/client-go/issues/374). If needed, users can tune the feature via the HTTP2_READ_IDLE_TIMEOUT_SECONDS and HTTP2_PING_TIMEOUT_SECONDS environment variables. The feature is disabled if HTTP2_READ_IDLE_TIMEOUT_SECONDS is set to 0. ([#100376](https://github.com/kubernetes/kubernetes/pull/100376),...

  property rights or otherwise. As a condition to exercising the
  rights and licenses granted hereunder, each Recipient hereby
  assumes sole responsibility to secure any other intellectual
  property rights needed, if any. For example, if a third party
  patent license is required to allow Recipient to Distribute the
  Program, it is Recipient's responsibility to acquire that license
  before distributing the Program.

## Changelog since v1.15.6

### Other notable changes

* Change GCP ILB firewall names to contain the "k8s-fw-" prefix like the rest of the firewall rules. This is needed for consistency and also for other components to identify the firewall rule as k8s/service-controller managed. ([#84622](https://github.com/kubernetes/kubernetes/pull/84622), [@prameshj](https://github.com/prameshj))

* On AttachDetachController node status update, do not retry when node doesn't exist but keep the node entry in cache. ([#50806](https://github.com/kubernetes/kubernetes/pull/50806), [@verult](https://github.com/verult))
* Prevent unneeded endpoint updates ([#50934](https://github.com/kubernetes/kubernetes/pull/50934), [@joelsmith](https://github.com/joelsmith))

be full of soup.

  `There's certainly too much pepper in that soup!' Alice said to
herself, as well as she could for sneezing.

  There was certainly too much of it in the air.  Even the
Duchess sneezed occasionally; and as for the baby, it was
sneezing and howling alternately without a moment's pause.  The
only things in the kitchen that did not sneeze, were the cook,

  - Speed up binpacking by reducing the number of PreFilter calls (call once per pod instead of #pods*#nodes times)
  - Speed up finding unneeded nodes by 5x+ in very large clusters by reducing the number of PreFilter calls
  - Expose `--max-nodes-total` as a metric
  - Errors in `IncreaseSize` changed from type `apiError` to `cloudProviderError`