- the IP allocator in v1.17.0 can return errors such as `the cluster IP <ip> for service <service-name> is not within the service CIDR <cidr>; please recreate` in the logs of the kube-apiserver. The cause is incorrect CIDR calculations if the service CIDR (`--service-cluster-ip-range`) is set to bits...

  - The `PersistentVolumeLabel` admission controller will be removed in a future release, when the out-of-tree versions of the GCE and AWS cloud providers move to GA. The cloud providers are marked alpha in version 1.9.

### OpenStack

To learn more about the feature, please refer to the official [seccomp tutorial](https://kubernetes.io/docs/tutorials/clusters/seccomp/#enable-the-use-of-runtimedefault-as-the-default-seccomp-profile-for-all-workloads).

### Quality of Service for memory resources

  - CinderLimits
  - EBSLimits
  - GCEPDLimits
  
  The NodeVolumeLimits plugin can handle the same functionality as the above plugins since the above volume types are migrated to CSI.
  Please remove those plugins and replace them with the NodeVolumeLimits plugin if you explicitly use those plugins in the scheduler config.

Rescheduler 0.3 shipped together with Kubernetes 1.6 understands the new taints and will clean up the old annotations, but Rescheduler 0.2 shipped together with Kubernetes 1.5 doesn't understand the new taints. In order to avoid eviction loop during 1.5->1.6 upgrade you need to either upgrade the master node atomically (for example by using `upgrade.sh` script for GCE) or ensure that the rescheduler is upgraded first, then the scheduler and then all the other master components.

* **Tolerations**...

e2e tests that require SSH may now be used against clusters that have nodes without external IP addresses by setting the environment variable `KUBE_SSH_BASTION` to the `host:port` of a machine that is allowed to SSH to those nodes.  The same private key that the test would use is used for the bastion host.  The test connects to the bastion and then tunnels another SSH connection to the node. ([#72286](https://github.com/kubernetes/kubernetes/pull/72286), [@smarterclayton](https://github.com/smarterclayton))...

- When using `kubectl replace` (or the equivalent API call) on a Service, the caller no longer needs to do a read-modify-write cycle to fetch the allocated values for `.spec.clusterIP` and `.spec.ports[].nodePort`.  Instead the API server will automatically carry these forward from the original object when the new object does not specify them. ([#104674](https://github.com/kubernetes/kubernetes/pull/104674), [@thockin](https://github.com/thockin))...

* The kube-apiserver discovery APIs (for example, `/apis`) return information about the API groups being served, and can change dynamically.
During server startup, prior to the server reporting healthy (via `/healthz`), not all API groups may be reported.
Wait for the server to report healthy (via `/healthz`) before depending on the information provided by the discovery APIs.

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 command line. Be sure to review the default values when migrating to using a config file. For example, the authz mode is set to "AlwaysAllow" if you rely on the command line, but defaults...

* `kubeadm join` is now blocking on the kubelet performing the TLS Bootstrap properly. Earlier, `kubeadm join` only did the discovery part and exited successfully without checking that the kubelet actually started properly and performed the TLS bootstrap correctly. Now, as kubeadm runs some post-join steps (for example, annotating the Node API object with the CRISocket), `kubeadm join` is now waiting for the kubelet to perform the TLS Bootstrap, and then uses that credential to perform...