k8s-and-chill/CLAUDE.md
Felix Wolf 85b8fec6b3 feat: replace secret-init Jobs with mittwald operator + cert-manager
Migrate ~180 LOC of openssl/kubectl init Jobs to declarative Secret
manifests reconciled by mittwald/kubernetes-secret-generator (random
strings, SSH keypair) and cert-manager Certificates (RSA private key +
self-signed CA chain). mittwald only fills empty fields, so existing
populated Secrets keep their current values across the migration.

Changes:

- New prototype kubernetes-secret-generator (chart 3.4.1, mittwald helm
  repo). Cluster-wide informer reconciler, no webhook -> cold-bootstrap
  safe via ArgoCD retries.
- New cert-manager selfsigned ClusterIssuer (in-cluster trust root).
  letsencrypt remains for public-DNS endpoints.
- forgejo: admin-secret Job replaced with a mittwald-annotated Secret
  (hex-encoded 24-char password). Deploy-key Job split: mittwald
  ssh-keypair Secret + slim Job that uploads pubkey to Forgejo and
  copies privkey into the argocd repo Secret.
- ocis: 13 Secrets / 16 random fields now mittwald-managed (UUIDs
  replaced with opaque random hex; ocis treats user-id as opaque). IDP
  RSA signing key, LDAP self-signed CA, and LDAP server cert produced
  by cert-manager. Per-Deployment ytt overlay remaps volume key paths
  (tls.crt -> ldap-ca.crt, tls.key -> private-key.pem, etc.) since the
  ocis chart mounts Secrets raw without items support. Old multi-secret
  s3-secret-job replaced with a slim external-secret precheck Job that
  only validates pre-created Hetzner S3/Storage Box credentials.
- Application sync-wave -10 on cert-manager and kubernetes-secret-
  generator so they install before consumers. ArgoCD selfHeal handles
  any residual races.

CLAUDE.md: remove the "all namespaces use privileged PodSecurity"
convention. Existing namespaces still carry the label and will be
audited separately.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-03 00:00:07 +02:00

85 lines
4.2 KiB
Markdown

# k8s-and-chill
## Project Overview
GitOps-managed Kubernetes cluster on Hetzner Cloud running Talos Linux. Uses [myks](https://github.com/mykso/myks) for Helm chart rendering with ytt overlays, targeting ArgoCD for continuous deployment.
## Cluster
- **3 Talos control-plane nodes** (CAX11 ARM64, Hetzner Cloud Nuremberg)
- Node IPs: `195.201.219.111`, `195.201.140.75`, `195.201.219.17`
- `allowSchedulingOnControlPlanes: true` (no dedicated workers)
## Domain & DNS
- **Domain**: `tr1ceracop.de` (registered at INWX)
- **DNS**: Managed at INWX with wildcard A record `*.tr1ceracop.de` pointing to node IPs
- **Forgejo**: `https://git.tr1ceracop.de`
- **ArgoCD**: `https://argocd.tr1ceracop.de`
## Deployed Applications
| App | Namespace | Notes |
|-----|-----------|-------|
| traefik | traefik | Ingress controller, DaemonSet with hostPort 80/443 |
| cert-manager | cert-manager | Let's Encrypt HTTP-01 via ClusterIssuer `letsencrypt` |
| forgejo | forgejo | Git server, SQLite, local-path PVC |
| argocd | argocd | GitOps controller |
| local-path-provisioner | local-path-storage | Default StorageClass, installed via upstream manifest |
## myks Structure
```
prototypes/ # Application templates (helm values + ytt overlays)
argocd/
traefik/
cert-manager/
forgejo/
envs/
env-data.ytt.yaml # Global ArgoCD config
_env/ # Shared overlays (annotations, secrets)
production/
env-data.ytt.yaml # App list for production
_apps/{app}/app-data.ytt.yaml # Per-app overrides
rendered/
envs/production/{app}/ # kubectl-ready manifests
argocd/production/ # ArgoCD Application resources
talos/
controlplane.yaml # Talos machine config
talosconfig # Talos client config
kubeconfig # Cluster kubeconfig
```
### Prototype Pattern
Each prototype follows this structure:
- `app-data.ytt.yaml` — namespace declaration
- `vendir/vendir-data.ytt.yaml` — chart name, version, repository URL
- `vendir/base.ytt.yaml` — vendir config template (identical across all)
- `helm/{chart}.yaml` — Helm values overrides
- `ytt/ns.ytt.yaml` — Namespace resource + namespace overlay on all resources
### Key Commands
```bash
myks render # Render all apps
myks render production <app> # Render single app
kubectl apply -f rendered/envs/production/<app>/ --server-side # Deploy
```
## Kubeconfig & Talos
- `KUBECONFIG` and `TALOSCONFIG` are already set in the user's shell environment. Do not set them in commands.
## Known Issues / TODOs
- **Namespace race condition**: First `kubectl apply` of a new app often fails because namespace isn't ready. Re-apply once.
- **Traefik DaemonSet updates**: Requires `updateStrategy.rollingUpdate.maxSurge: 0` because hostPort conflicts prevent surge.
- **Forgejo Ingress API version**: Chart renders `extensions/v1beta1`, fixed via `ytt/ingress-fix.ytt.yaml` overlay to `networking.k8s.io/v1`.
- **ArgoCD**: Fully wired to Forgejo via App of Apps. Root Application in `default` project syncs `rendered/argocd/production/`. Deploy key provisioned automatically by `argocd-deploy-key-init` Job in forgejo namespace.
## Container Images
- **Never use bitnami images.** Use `alpine/k8s` or plain `alpine` for utility Jobs instead.
## Secrets
- **Never commit secrets to git.** This is a public repository.
- **All secrets must be generated in-cluster** using init Jobs (ArgoCD PreSync hooks) that create secrets if they don't already exist. See `prototypes/ocis/ytt/s3-secret-job.ytt.yaml` for the pattern.
- **External secrets** (e.g. S3 credentials) that cannot be generated must be created manually in the cluster before deploying. The init Job should validate their existence and fail fast if missing.
- When adding a new application that uses a Helm chart generating secrets, configure all `secretRefs` to point to pre-created secret names and use an init Job to generate them.
- Known external secrets (not in git, created manually):
- `ocis/ocis-s3-credentials` — Hetzner S3 access key and secret key
- `ocis/ocis-storagebox-credentials` — Hetzner Storage Box host, user, and SSH private key (for S3 backup to Helsinki)
- `cert-manager/letsencrypt-account-key` — ACME account key (auto-generated by cert-manager)