Audited the running cluster and fixed all .md files: - Node info: Fedora 43, Lima (not OrbStack), worker IP 10.0.1.58 - Networking: fixed public/internal hostname tables, all *.dog internals - Storage: removed Longhorn refs (not deployed), documented hostPath/local-path - Services: moved Seerr to media chart, utils is Zerobyte only - Bootstrap: reordered steps, MetalLB/traefik-internal as manual pre-deploy - Headlamp.md/MetalLB.md: added context and explanations Made-with: Cursor
1.7 KiB
Executable File
NAME STATUS ROLES AGE VERSION INTERNAL-IP EXTERNAL-IP OS-IMAGE KERNEL-VERSION CONTAINER-RUNTIME localhost.localdomain Ready control-plane v1.34.6+k3s1 10.0.1.2 Fedora Linux 43 (Server Edition) 6.17.1-300.fc43.x86_64 containerd://2.2.2-bd1.34 lima-mac-worker Ready v1.34.6+k3s1 10.0.1.58 Ubuntu 25.10 6.17.0-22-generic containerd://2.2.2-bd1.34
The mac-worker is running inside a Lima VM on macOS.
I have a DNS rewrite pointing *.internal to 10.0.1.250 which is traefik-internal.
/dogstore/ is a NFS path that's available to all nodes
secrets are managed by sops
Load balancers
Two LB implementations coexist: k3s klipper (servicelb) and MetalLB. They are
separated by loadBalancerClass so they don't conflict.
- klipper handles services with NO
loadBalancerClass. It creates svclb DaemonSet pods that bind host ports directly on every node. - MetalLB handles services with
loadBalancerClass: metallb. Its pool hasautoAssign: false, so it only assigns IPs to services that explicitly request a pool via themetallb.io/address-poolannotation.
| Service | loadBalancerClass | LB | External IPs |
|---|---|---|---|
| traefik | (none) | klipper | 10.0.1.2, 10.0.1.58 |
| traefik-internal | metallb | MetalLB | 10.0.1.250 |
loadBalancerClass is immutable on k8s Services. Changing it requires deleting
the Service first, then redeploying (kubectl delete svc … && helm upgrade).