### Client
* Use Prometheus instrumentation conventions ([#36704](https://github.com/kubernetes/kubernetes/pull/36704), [@fabxc](https://github.com/fabxc))
* Clients now use the `?watch=true` parameter to make watch API calls, instead of the `/watch/` path prefix ([#41722](https://github.com/kubernetes/kubernetes/pull/41722), [@liggitt](https://github.com/liggitt))

## Before Upgrading

Consider the following changes, limitations, and guidelines before you upgrade:

### **API Machinery**

*   The admission API, which is used when the API server calls admission control webhooks, is moved from `admission.v1alpha1` to `admission.v1beta1`. You must **delete any existing webhooks before you upgrade** your cluster, and update them to use the latest API. This change is not backward compatible.

- Avoid GCE API calls when initializing GCE CloudProvider for Kubelets. ([#90218](https://github.com/kubernetes/kubernetes/pull/90218), [@wojtek-t](https://github.com/wojtek-t)) [SIG Cloud Provider and Scalability]

 - Reduced frequency of DescribeVolumes calls of AWS API when attaching/detaching a volume.

* GCE provider: Create TargetPool with 200 instances, then update with rest ([#27865](https://github.com/kubernetes/kubernetes/pull/27865), [@zmerlynn](https://github.com/zmerlynn))
* GCE provider: Limit Filter calls to regexps rather than large blobs ([#27741](https://github.com/kubernetes/kubernetes/pull/27741), [@zmerlynn](https://github.com/zmerlynn))

	madminSysConfig        = grid.NewJSONPool[madmin.SysConfig]()
	madminSysErrors        = grid.NewJSONPool[madmin.SysErrors]()
	madminSysServices      = grid.NewJSONPool[madmin.SysServices]()

	// Request -> Response RPC calls
	deleteBucketMetadataRPC        = grid.NewSingleHandler[*grid.MSS, grid.NoPayload](grid.HandlerDeleteBucketMetadata, grid.NewMSS, grid.NewNoPayload).IgnoreNilConn()

* fix device name change issue for azure disk: add remount logic ([#57953](https://github.com/kubernetes/kubernetes/pull/57953), [@andyzhangx](https://github.com/andyzhangx))
* falls back to parse Docker runtime version as generic if not semver ([#54040](https://github.com/kubernetes/kubernetes/pull/54040), [@dixudx](https://github.com/dixudx))

  std::vector<std::pair<string, string>> grads = GetGradDefs(gdef);
  ASSERT_EQ(1, grads.size());
  ASSERT_EQ(func_name_, grads[0].first);
  ASSERT_EQ("MyGrad", grads[0].second);

  // These calls must be noops
  TF_GraphCopyFunction(host_graph_, func_, grad_func, s_);
  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);
  TF_GraphCopyFunction(host_graph_, func_, nullptr, s_);

- Fix the regression that introduced 34s timeout for DELETECOLLECTION calls ([#115481](https://github.com/kubernetes/kubernetes/pull/115481), [@tkashem](https://github.com/tkashem)) [SIG API Machinery]

          .build(),
      )
    executeSynchronously(request).assertFailure("$dns returned no addresses for android.com")
    dns.assertRequests("android.com")
  }

  /** We had a bug where failed HTTP/2 calls could break the entire connection.  */
  @Flaky
  @Test
  fun failingCallsDoNotInterfereWithConnection() {
    enableProtocol(Protocol.HTTP_2)
    server.enqueue(MockResponse(body = "Response 1"))