New
- gVisor (runsc) as an additional container runtime. The new
addons.gvisorsetting, off by default, installsrunscon every cloud node, registers it as a containerd runtime and creates a cluster-scopedgvisorRuntimeClass, so a pod that setsruntimeClassName: gvisorruns in a user-space kernel instead of on the host kernel. - @rightsum- Debian and Ubuntu only. Setup stops with an explicit error on an image that has no
apt-get, and again if runsc is still missing after the install, rather than finishing with a RuntimeClass that points at a runtime the node never loaded. addons.gvisor.versionpins the runsc build, default20261116.0. The gVisor apt repository has a single rolling component, so an unpinned install would upgrade the runtime on every node at everycreate.- Enabling it on an existing cluster writes the containerd template, and
createrestarts the k3s agent on each cloud worker whose file changed. A master restarts only withschedule_workloads_on_masters, since a tainted master never runs these pods. Autoscaled nodes pick the runtime up when they launch. External nodes that are already initialized are not re-provisioned, and the tool warns instead. - Covered in the cluster creation guide, the addons list, the upgrade notes, and a troubleshooting entry for a pod that asks for the runtime on a node that does not have it.
- Debian and Ubuntu only. Setup stops with an explicit error on an image that has no
Improvements
- The gVisor guide now says which containerd template format is written and why it works on the containerd versions k3s ships.
- Troubleshooting covers gVisor pods stuck in
ContainerCreating, from a node that predates the addon to an external node that skips the setup script.
Upgrading
No action is required. The addon is off by default and a cluster that does not enable it gets byte identical setup scripts: the gVisor block is left out of the render entirely, which was checked against a real config with the addon off before the release.
Enabling it on an existing cluster restarts the k3s agent once on each cloud worker whose containerd template changed, which does not interrupt running pods, and leaves the control plane alone unless schedule_workloads_on_masters is true. Details are in the upgrade notes.
No addon versions changed in this release: the cloud controller manager, CSI driver, system upgrade controller and cluster-autoscaler pins are the same as in v2.14.0.
If hetzner-k3s is useful to you or your company, please consider sponsoring its development.
Sponsorship helps ensure continued maintenance and new features. Thank you to our current sponsors! 🙏