This release updates Kubernetes to v1.22.3, and fixes a number of issues.
For more details on what's new, see the Kubernetes release notes.
- There is an issue that may cause unexpected behavior when removing servers running embedded etcd from the cluster. When removing servers with etcd from the cluster, ensure that the k3s service is stopped before using
kubectl deleteto remove the node from the Kubernetes and etcd cluster. You should do this one node at a time to ensure that your cluster does not lose quorum. Failure to follow this process may result in K3s on the deleted node crashing, restarting, and rejoining the cluster.
- There is an issue that affects k3s when deploying on dual-stack mode on hosts with a kernel version older than 5.1, the IPv6 external-ip exposed by the service LB does not work correctly.
Changes since v1.22.2+k3s2:
- Updated Kubernetes to v1.22.3 (#4348)
- Added dual-stack support LB controller (#4160)
- New aliases for k3s etcd-snapshot s3 commands to match the flags used with k3s server (#4170) (#4282)
- Set skip SSL verification for etcd snapshots if skip ssl verify flag is passed (#4103)
- Display cluster tls error only in debug mode (#4200)
- Ensured that node address labels are set before initializing the node with the integrated cloud provider (#4222)
--timeoutoption to etcd-snapshot command to allow increasing the time allowed for snapshot uploads to complete (#4230)
- K3s should now reliably exit when core Kubernetes components (apiserver, controller-manager, etc) experience fatal errors (#4241)
- K3s now waits until etcd is actually able to start before joining a new members to the embedded etcd clusters (#4245)
- Fixed an issue that caused k3s to not be restarted by systemd when started with the
- Updated flannel to v0.15.1 (#4265)
- Updated helm-controller and klipper-helm to time out the
helm_v2 ls --allthat is done to check for deprecated helm chart versions (#4291)
- Fixed a problem in etcd cluster-reset by updating the peer addresses (#4305)
- When using the integrated cloud provider, ensure that the scheduler doesn't start before at least one node in the cluster has been initialized (#4361)
Embedded Component Versions
As always, we welcome and appreciate feedback from our community of users. Please feel free to: