Cilium clusters now validate their network at the end of create, and the MTU is a configuration field, so a setup that would blackhole large packets fails at creation time instead of in production.
New
networking.cni.cilium.mtuandnetworking.cni.cilium.underlay_mtu. Leave both unset to keep Cilium's auto-detection. Whenmtuis set,createrejects it before creating any server if it does not fit the underlay, or if it leaves pods below the MTU floor. - by @Yiidiir- A cross-node pod network check at the end of
create: two short-lived pods on two Ready nodes, a small ping between them, then a ping as large as the pod MTU with the don't-fragment bit set. - by @Yiidiir networking.cni.cilium.connectivity_check, withenabled,imageandtimeout_seconds, for image mirrors, air-gapped clusters, or skipping the check. - by @Yiidiir- A troubleshooting entry for the new failure message and what each part of it means. - by @Yiidiir
Improvements
- MTU validation follows the Cilium chart in use: WireGuard costs 80 bytes up to chart v1.19.0 and 95 bytes from v1.19.5, so the expected pod MTU matches what Cilium actually programs.
- On a dual stack pod network the pod MTU floor is 1280 rather than 1230, because pods lose IPv6 below that.
- Clusters that set no MTU render exactly the same Helm values as before.
- The check compares the pod MTU with the configured one in both directions, because changing
mtudoes not restart the Cilium agents and pods can keep the previous size. - Specs cover the validator, the Helm values parsing, the overhead table, dual stack detection, and the generated check script.
Upgrading
No configuration change is required, and the bundled addon pins are unchanged since v2.15.0.
Cilium clusters run the network check on the next create, including a re-run against a cluster that already exists. It needs at least two Ready nodes and pulls nicolaka/netshoot:v0.16 on each test node, so air-gapped clusters or clusters that pull only through a registry mirror should point networking.cni.cilium.connectivity_check.image at a reachable copy, or disable the check. Clusters with a single node skip it and print a warning.
Changing mtu on a cluster that already exists needs kubectl -n kube-system rollout restart ds cilium before the next create, since the Helm upgrade does not restart the agents.
The important upgrade notes cover both cases.
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! 🙏