| apps | ||
| argocd-install | ||
| bootstrap | ||
| cluster | ||
| infra/traefik | ||
| scripts | ||
| README.md | ||
argocd
GitOps for the homelab k3s cluster (k3s-design.md). Argo CD watches this repo
on Forgejo and deploys whatever is on main.
Layout
bootstrap/root.yaml the only thing applied by hand: the "root" Application -> cluster/
cluster/ cluster-level Argo objects (app-of-apps)
argocd.yaml Argo CD manages its own install from argocd-install/
apps.yaml ApplicationSet: one Application per folder in apps/
sealed-secrets.yaml Sealed Secrets controller (kube-system)
longhorn.yaml Longhorn storage (longhorn-system)
traefik-config.yaml tweaks to k3s's Traefik, from infra/traefik/
argocd-install/ Argo CD install (upstream install.yaml pinned + small patches)
infra/ plain manifests used by cluster/ Applications
apps/<name>/ one app per folder, deployed to namespace <name>
scripts/ prep-worker.sh (Longhorn prerequisites), seal-secrets.sh
Argo reads the repo at http://10.0.0.14:3000/un-oeil/argocd.git (Forgejo's
internal address). The repo is public, so no credentials are needed.
Apps
| App | URL | Data | Database |
|---|---|---|---|
| homepage | home.un-oeil.fr | config in git | none |
| ntfy | ntfy.un-oeil.fr | 2 Gi | SQLite |
| wikijs | wiki.un-oeil.fr | none | Postgres |
| paperless | paperless.un-oeil.fr | 2 + 50 + 5 Gi | Postgres + Valkey |
| linkwarden | links.un-oeil.fr | 20 Gi | Postgres |
| firefly | firefly.un-oeil.fr | 2 Gi | Postgres |
| hedgedoc | md.un-oeil.fr | 5 Gi | Postgres |
| freshrss | rss.un-oeil.fr | 2 Gi | SQLite |
| vikunja | tasks.un-oeil.fr | 5 Gi | SQLite |
| n8n | n8n.un-oeil.fr | 2 Gi | SQLite |
| searxng | search.un-oeil.fr | none | none |
| privatebin | paste.un-oeil.fr | 1 Gi | files |
| stirling-pdf | pdf.un-oeil.fr | none | none |
| it-tools | tools.un-oeil.fr | none | none |
| postgres | (internal) postgres.postgres.svc |
10 Gi | shared by the 5 above |
Every volume is Longhorn, 2 replicas, on the workers' sdb. Total requests are
about 3 GiB RAM and 110 GiB of volumes (×2 replicas = 220 of 600 GiB).
First rollout (in this order)
Over WireGuard, with KUBECONFIG=~/.kube/k3s-homelab.yaml.
1. Prepare the workers for Longhorn. On each of 10.0.0.141, .142, .143:
ssh un-oeil@10.0.0.141 'sudo bash -s' < scripts/prep-worker.sh
It formats /dev/sdb, which is the worker's own empty 200 GB disk (the zvol
hdd-mirror/vm-disks/vm-14x-disk-0 on the host). It is not
/hdd-mirror/vzdump-backups or hdd-mirror/restic-repo. The script stops if
sdb isn't 200 GB or already holds anything. It also installs open-iscsi and
nfs-common.
2. Push the cluster layer only:
git add cluster/ infra/ scripts/ README.md
git commit -m "Add Sealed Secrets, Longhorn, Traefik config"
git push
Wait until both are healthy:
kubectl -n kube-system rollout status deploy/sealed-secrets-controller
kubectl -n longhorn-system get pods # all Running
kubectl get nodes.longhorn.io -n longhorn-system # 3 workers, Ready
3. Back up the Sealed Secrets key. Without it, nobody can decrypt the secrets in this repo after a cluster rebuild. Keep it out of git, e.g. in Vaultwarden:
kubectl -n kube-system get secret -l sealedsecrets.bitnami.com/sealed-secrets-key -o yaml > sealed-secrets-key.yaml
4. Generate and seal the app secrets, then push the apps:
scripts/seal-secrets.sh
git add apps/
git commit -m "Add apps"
git push
Postgres needs to start before the five apps that use it. Those apps restart a few times until it's ready, which is normal.
5. Caddy (LXC 100). One block for every app. Caddy keeps the Host header,
and Traefik routes on it:
home.un-oeil.fr, ntfy.un-oeil.fr, wiki.un-oeil.fr, paperless.un-oeil.fr,
links.un-oeil.fr, firefly.un-oeil.fr, md.un-oeil.fr, rss.un-oeil.fr,
tasks.un-oeil.fr, n8n.un-oeil.fr, search.un-oeil.fr, paste.un-oeil.fr,
pdf.un-oeil.fr, tools.un-oeil.fr {
reverse_proxy 10.0.0.141:80 10.0.0.142:80 10.0.0.143:80 {
lb_policy round_robin
fail_duration 30s
# ntfy and HedgeDoc stream
flush_interval -1
}
}
Then env $(cat /etc/caddy/infomaniak.env) caddy validate --config /etc/caddy/Caddyfile && systemctl reload caddy.
Expect 14 certificate obtained lines in the journal.
6. DNS. Add the 14 A records → 10.0.0.10 at Infomaniak (zone file in
../infomaniak-zone-un-oeil.fr.txt).
After the first deploy
- Paperless: user
admin. Read the password withkubectl -n paperless get secret paperless-secrets -o jsonpath='{.data.PAPERLESS_ADMIN_PASSWORD}' | base64 -d, then change it. - Wiki.js, FreshRSS, n8n, Firefly, Linkwarden, HedgeDoc, Vikunja: the first visit creates the admin or owner account. Do it right after the deploy.
- Vikunja: once your account exists, set
VIKUNJA_SERVICE_ENABLEREGISTRATIONto"false"inapps/vikunja/deployment.yaml. - HedgeDoc: same, with
CMD_ALLOW_EMAIL_REGISTER: "false". - ntfy: nobody can use it until you create users:
kubectl -n ntfy exec deploy/ntfy -- ntfy user add --role=admin un-oeil kubectl -n ntfy exec deploy/ntfy -- ntfy token add un-oeil # for Uptime Kuma / scripts - Uptime Kuma: add the 14 URLs (
uptime-kuma-monitors.md§3), plus a Push monitor for the Postgres backup below.
Postgres backups
apps/postgres/backup-cronjob.yaml streams pg_dumpall into restic on LXC 105
every night at 02:45, keeps 7 daily / 4 weekly / 3 monthly, then pings Uptime
Kuma. It's suspended until it has credentials:
- On LXC 105, add a rest-server user
k3s-postgres, named after its repo like the existing ones (--private-reposlimits each user to its own path):htpasswd -B /etc/restic-htpasswd k3s-postgres && systemctl restart rest-server - Create a Push monitor "Backup – Postgres (k3s)" in Kuma, interval 90000 s.
- Seal the credentials. The script adds them to the kustomization:
RESTIC_REPOSITORY='rest:http://k3s-postgres:<pw>@10.0.0.15:8000/k3s-postgres' \ RESTIC_PASSWORD='<new repo password, keep it in Vaultwarden>' \ KUMA_PUSH_URL='https://status.un-oeil.fr/api/push/<token>?status=up&msg=OK' \ scripts/seal-secrets.sh - Set
suspend: falsein the CronJob, push, then test it once:kubectl -n postgres create job --from=cronjob/postgres-backup test-backup.
Not covered: the files on Longhorn volumes (Paperless documents, Linkwarden
archives, uploads). The workers' sdb disks are excluded from vzdump
(backup=0), and Longhorn has no backup target yet. Longhorn's backup target
can be an NFS export or S3 bucket, e.g. on the restic LXC. Until then, those
files survive a node failure (2 replicas) but not the loss of the pool.
Launching an app
- Create
apps/<name>/. Its contents can be:- plain manifests, or
- a
kustomization.yaml(Helm charts work too viahelmCharts:), or - a
Chart.yaml(umbrella chart with the real chart as a dependency +values.yaml).
- Volumes:
storageClassName: longhorn, andstrategy: Recreateon the Deployment (a Longhorn volume attaches to one node at a time). PutenableServiceLinks: falsein the pod spec: otherwise a Service named like the app injects<APP>_PORT=tcp://..., which some apps (Paperless) read as their own setting. - Secrets: add a
sealline toscripts/seal-secrets.sh, or runkubesealby hand. Never commit a plain Secret. - A Postgres database: create it by hand, because the init script only runs on
the first start:
kubectl -n postgres exec -it postgres-0 -- psql -U postgres -c "CREATE ROLE x LOGIN PASSWORD '...'" -c "CREATE DATABASE x OWNER x" - Commit and push. Argo picks it up within ~3 min and creates namespace
<name>. - Ingress host
<name>.un-oeil.fr, add the name to the Caddy block above, and an A record →10.0.0.10.
Removing a folder deletes the app and its resources, PVCs included.
Longhorn, Sealed Secrets and Argo CD itself have prune: false, so a bad
commit can't delete them.
Bootstrap (once, already done 2026-10-09)
kubectl apply --server-side -k argocd-install/
kubectl -n argocd rollout status deploy/argocd-server
kubectl apply -f bootstrap/root.yaml
Argo CD UI: https://argocd.un-oeil.fr through Caddy, behind Argo's own login.
Caddy can't restrict it to the admin laptop: the WireGuard hub masquerades
every peer as 10.0.0.250, so remote_ip never sees 10.10.10.2.
Without Caddy: kubectl -n argocd port-forward svc/argocd-server 8080:80.
Longhorn UI: kubectl -n longhorn-system port-forward svc/longhorn-frontend 8081:80
(or the same Caddy pattern with a longhorn.un-oeil.fr Ingress).
Before relying on it
- Image versions were current on 2026-10-11. Several are new major versions (Paperless 3, Vikunja 2, n8n 2, Homepage 2, Stirling-PDF 3), so check their release notes for renamed settings if one misbehaves.
- Traefik trusts
X-Forwarded-Forfrom10.0.0.0/24and the pod network, so apps see real client IPs through Caddy. Someone on WireGuard who hits Traefik directly could fake that header. That's acceptable for a private network. - Not there yet: Forgejo webhook → Argo for instant syncs, Longhorn backup
target, kube-vip on
.139.