- Add a new `distribute-cpus-across-numa` option to the static `CPUManager` policy. When enabled, this will trigger the `CPUManager` to evenly distribute CPUs across NUMA nodes in cases where more than one NUMA node is required to satisfy the allocation. ([#105631](https://github.com/kubernetes/kubernetes/pull/105631), [@klueska](https://github.com/klueska))

- 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))

- Promoted the Topology Manager policy option `max-allowable-numa-nodes` to GA version. ([#134614](https://github.com/kubernetes/kubernetes/pull/134614), [@ffromani](https://github.com/ffromani))

## Changelog since v1.21.5

## Changes by Kind

### API Change

- Kube-apiserver: Fixes handling of CRD schemas containing literal null values in enums (#104989, @liggitt) [SIG API Machinery, Apps and Network]

### Feature

- Kubernetes is now built with Golang 1.16.9 (#105672, @cpanato) [SIG Cloud Provider, Instrumentation, Release and Testing]

- Added a `--topology-manager-policy-options` flag to the kubelet to support fine tuning the topology manager policies. The first policy option, `prefer-closest-numa-nodes`, allows these policies to favor sets of NUMA nodes with shorter distance between nodes when making admission decisions. ([#112914](https://github.com/kubernetes/kubernetes/pull/112914), [@PiotrProkop](https://github.com/PiotrProkop))

* 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))...

- Omit enum declarations from the static openapi file captured at https://git.k8s.io/kubernetes/api/openapi-spec. This file is used to generate API clients, and use of enums in those generated clients (rather than strings) can break forward compatibility with additional future values in those fields. See https://issue.k8s.io/109177 for details. ([#109178](https://github.com/kubernetes/kubernetes/pull/109178), [@liggit...

- Added metrics to capture CPU distribution across NUMA nodes ([#130491](https://github.com/kubernetes/kubernetes/pull/130491), [@swatisehgal](https://github.com/swatisehgal)) [SIG Node and Testing]

- CEL cost estimator no longer treats enums as unbounded strings when determining its length. Instead, the length is set to the longest possible enum value. ([#121085](https://github.com/kubernetes/kubernetes/pull/121085), [@jiahuif](https://github.com/jiahuif)) [SIG API Machinery]

- Add --for=create option to kubectl wait ([#125868](https://github.com/kubernetes/kubernetes/pull/125868), [@soltysh](https://github.com/soltysh)) [SIG CLI and Testing]
- Add a TopologyManager policy option: max-allowable-numa-nodes to configures maxAllowableNUMANodes for kubelet. ([#124148](https://github.com/kubernetes/kubernetes/pull/124148), [@cyclinder](https://github.com/cyclinder)) [SIG Node and Testing]