After the connection closed(received GOAWAY), the client's other in-flight requests won't be affected, and the client will reconnect. 
  The flag min value is 0 (off), max is .02 (1/50 requests); .001 (1/1000) is a recommended starting point.

  * Kubernetes now shares a single PID namespace among all containers in a pod when running with docker >= 1.13.1. This means processes can now signal processes in other containers in a pod, but it also means that the `kubectl exec {pod} kill 1` pattern will cause the Pod to be restarted rather than a single container. ([#45236](https://github.com/kubernetes/kubernetes/pull/45236), [@verb](https://github.com/verb))

- Fixed a bug in the single-numa-policy of the TopologyManager. Previously, best-effort pods would result in a terminated state with a TopologyAffinity error. Now they will run as expected. ([#83777](https://github.com/kubernetes/kubernetes/pull/83777), [@lmdaly](https://github.com/lmdaly))

  It changes the dual-stack API wrt Service from a single ipFamily field to 3
  fields: ipFamilyPolicy (SingleStack, PreferDualStack, RequireDualStack),
  ipFamilies (a list of families assigned), and clusterIPs (inclusive of
  clusterIP).  Most users do not need to set anything at all, defaulting will
  handle it for them.  Services are single-stack unless the user asks for

* Added anti-affinity to kube-dns pods. Otherwise the "no single point of failure" setting doesn't actually work (a single node failure can still take down the entire cluster). ([#57683](https://github.com/kubernetes/kubernetes/pull/57683), [@vainu-arto](https://github.com/vainu-arto))

### API-Machinery

  
  Multi vCenter support is deprecated as of v1.21. If you have a Kubernetes cluster spanning across multiple vCenter servers, please consider moving all k8s nodes to a single vCenter Server. vSphere CSI Driver does not support Kubernetes deployment spanning across multiple vCenter servers.
  

- Promoted `RelaxedDNSSearchValidation` to beta, allowing for Pod search domains to be a single dot "." or contain an underscore "_". ([#130128](https://github.com/kubernetes/kubernetes/pull/130128), [@adrianmoisey](https://github.com/adrianmoisey)) [SIG Apps and Network]

- For GCE cluster provider, fix bug of not being able to create internal type load balancer for clusters with more than 1000 nodes in a single zone. ([#89902](https://github.com/kubernetes/kubernetes/pull/89902), [@wojtek-t](https://github.com/wojtek-t)) [SIG Cloud Provider, Network and Scalability]

*   Custom resources served by CustomResourceDefinition objects now support field selectors for `metadata.name` and `metadata.namespace`. Also fixed an issue with watching a single object; earlier versions could watch only a collection, and so a watch on an instance would fail.  ([#53345](https://github.com/kubernetes/kubernetes/pull/53345),[ @ncdc](https://github.com/ncdc))

#### **Other**

- Added `singleProcessOOMKill` flag to the kubelet configuration. Setting that to true enable single process OOM killing in cgroups v2. In this mode, if a single process is OOM killed within a container, the remaining processes will not be OOM killed. ([#126096](https://github.com/kubernetes/kubernetes/pull/126096), [@utam0k](https://github.com/utam0k)) [SIG API Machinery, Node,...