Sixteen reported issues, plus a handful of things found while working through them.
New configuration
networking.public_network.use_hetzner_firewall: falsecreates no Hetzner Cloud firewall at all. It is attached by label selector, so this matters for projects near the firewall and rule limits. The in-node local firewall is a separate setting.addons.coredns.enabled: truelets the tool manage CoreDNS itself, with configurableresourcesandreplicas. Off by default, so k3s keeps its own. The object names, thek8s-app: kube-dnslabel and thekube-dnsClusterIP are preserved, since the node-local cache forwards to that label.reconcile_node_pools: truemakescreatedrain and delete workers that the configuration no longer renders, so the file can be the source of truth. It lists what it would remove and asks first,--forceapplies without asking.networking.ssh.private_keyandpublic_keytake the key pair from the environment, so unattended runs do not have to write key files first.HCLOUD_TOKENnow takes precedence overhetzner_tokenin the config, so a token can stay out of the file. One line is printed if both are set and differ.addons.cluster_autoscaler.new_pod_scale_up_delaysets the flag of the same name.storage_class.local_path_non_default: truekeepslocal-pathas a non-default class instead of stripping it, for clusters where something else should be the default.- IPv6-only clusters work: with
public_network.ipv4: falsethe node address is used and bracketed in URLs, and the API certificate carries an IPv6 SAN.
Fixes
- Worker pool names are validated.
_fails, a leading digit is fine, since the name only ever appears inside a server name. - Both install scripts run under
sh, which is what cloud-init uses, and a failed k3s installer download is a failed install rather than a silent success. k3s-agentis restarted only ifregistries.yamlactually changed, not on everycreate.- The Hetzner components restart if the API token changes. They read it once, at pod start, so a rotated token previously needed a manual rollout.
- The rendered
firewall.shno longer embeds the read-write project token as a world-readable literal. It moved to/etc/hetzner-k3s/hetzner-ips-token, mode 0600, and the optional read-onlyhetzner_ips_tokencan be used for the lookup. - An explicit
falsefor the Hubble toggles is honoured.false || trueistruein Crystal, so the option was accepted and then ignored. - The autoscaler ClusterRole gets the
resource.k8s.iopermissions it needs since 1.35, where DRA is on and locked. - The autoscaler manifest is shipped with the binary at the release matching the default tag instead of following upstream
master, which brokecreateon every upstream manifest change. - The instance readiness waits are bounded and report what failed instead of looping with no output. The current IP lookup got timeouts too.
- The
local-pathannotation lookup uses the right jsonpath, so it no longer reports every patch as overwritten and abortscreate. - Pool reconciliation finds autoscaled nodes by their
hcloud/node-groupserver label, leaves external nodes alone unless their pool is gone, and refuses to remove anything if it cannot list the cluster.
Upgrading
docs/Important_upgrade_notes.md covers the two behaviour changes worth knowing: the HCLOUD_TOKEN precedence, and the one-off master restart on IPv6-only clusters. Existing configuration files need no edit.
Thanks
To everyone who reported and reproduced these: @DeprecatedLuke, @imantaba, @arfath-linklet, @deubert-it, @KristijanL, @reinhardmunz, @krystian-blaszczyk, @mertcangokgoz, @elderapo, @lopcode, @an-tex, @panroman14, @viniciusdc, @MengLinMaker, @MBR-0001 and @recepdmr.
Full changelog: v2.11.0...v2.12.0
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!