- When the go-client reflector relists, the ResourceVersion list option is set to the reflector's latest synced resource version to ensure the reflector does not "go back in time" and reprocess events older than it has already processed. If the server responds with an HTTP 410 (Gone) status code response, the relist falls back to using `resourceVersion=""`. ([#83520](https://github.com/kubernetes/kubernetes/pull/83520), [@jpbetz](https://github.com/jpbetz))

([#44744](https://github.com/kubernetes/kubernetes/pull/44744), [@yguo0905](https://github.com/yguo0905))

  * 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](...

for COS images on GCE has changed to `/home/kubernetes/flexvolume`, rather than `/etc/srv/kubernetes/kubelet-plugins/volume/exec`. Existing Flexvolume installations in clusters using COS images must be moved to the new directory, and installation processes must be updated with the new path. ([#58171](https://github.com/kubernetes/kubernetes/pull/58171), [@verult](https://github.com/verult))

* Default values differ between the Kubelet's componentconfig (config file) API and the Kubelet's...

a pod only applied to CPU resources (such as `cpu_shares`). The Kubernetes cgroup manager uses `memory.limit_in_bytes` in v1 cgroups to limit the memory capacity for a container, and uses `oom_scores` to recommend an order for killing container processes if an out-of-memory event occurs. This implementation has shortcomings: for `Guaranteed` pods, memory can not be fully reserved, and the page cache is at risk of being recycled. For `Burstable` pods, overcommitting memory (setting `request` less...

- 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, Testing and Windows]

  For more information, see [#47131](https://github.com/kubernetes/kubernetes/issues/47131)

*   If a namespaced resource is owned by a cluster scoped resource, and the namespaced dependent is processed before the cluster scoped owner has ever been observed by the garbage collector, the dependent will be erroneously deleted.

  For more information, see [#54940](https://github.com/kubernetes/kubernetes/issues/54940)

          ([#275](https://github.com/containerd/containerd/issues/275))

        - Additional memory overhead per container.
          ([#21737](https://github.com/kubernetes/kubernetes/pull/21737))

        - Processes may be leaked when Docker is repeatedly terminated in a short
          time frame.
          ([#41450](https://github.com/kubernetes/kubernetes/issues/41450))

- Added `kubelet` Topology Manager metrics to track admission requests processed and occured admission errors. ([#115137](https://github.com/kubernetes/kubernetes/pull/115137), [@swatisehgal](https://github.com/swatisehgal))

- GCE/Windows: When the service cannot be stopped Stackdriver logging processes are now force killed ([#77378](https://github.com/kubernetes/kubernetes/pull/77378), [@yujuhong](https://github.com/yujuhong))

### SIG-cloud-provider

This is the first Kubernetes release for this SIG! In v1.12, SIG Cloud Provider focused on building the processes and infrastructure to better support existing and new cloud providers. Some of these initiatives (many of which are still in progress) are: