bluejay-infra

Infrastructure manifests for ArgoCD. An ApplicationSet in argocd namespace watches the apps/* directories in this repo and creates one Application per subdir (prefixed infra-<name>).

Adding a new service to the cluster

Follow these steps in order. Step 1 must run before step 3 — if you skip it, cert-manager HTTP-01 will silently fail for ~2h per cert (exponential backoff) until someone diagnoses the DNS.

1. Create or verify the FlowerCore.DNS A record (REQUIRED for current HTTP-01 manifests)

step-ca (the ACME CA on noc1) runs in a Podman container with host networking. Its container resolver uses pfSense Unbound (10.0.56.1), not cluster CoreDNS. So even though CoreDNS has a wildcard *.iamworkin.lan → 10.0.56.200 for in-cluster lookups, step-ca cannot see it. Every new public hostname needs an explicit pfSense host override.

The management path is now FlowerCore.DNS, not FlowerCore.Notes/scripts/pfsense-add-dns-overrides.py. Add or verify the public A record there before you apply the manifest:

curl -sk https://dns.iamworkin.lan/api/v1/servers
# Find the pfSense serverId, then create the record using the host label only.
# Example: for foo.iamworkin.lan, use "name":"foo".

curl -sk -X POST https://dns.iamworkin.lan/api/v1/servers/<serverId>/zones/iamworkin.lan/records \
  -H "Content-Type: application/json" \
  -d '{"name":"<yourservice>","type":"A","data":"10.0.56.200","ttl":300}'

Verify all referenced iamworkin.lan hosts resolve (run from anywhere on LAN):

python scripts/check-pfsense-dns.py
# Historical filename retained. The script now calls
# https://dns.iamworkin.lan/api/v1/zones/iamworkin.lan/resolve-preflight
# for every Certificate dnsName and Traefik Host(...) rule it finds.

python scripts/check-pfsense-dns.py --live
# Optional stronger pass when kubectl access is available; also checks
# live-cluster Certificates and IngressRoutes for drift outside manifests.

Symptom if you skip this: the Certificate resource stays Ready: False with status.reason: unexpected non-ACME API error: context deadline exceeded. Recovery requires kubectl -n <ns> delete order <order-name> after adding the DNS to bypass cert-manager's backoff.

2. Create the app manifest

Create apps/<name>/<name>.yaml containing the Namespace, Deployment, Service, Certificate, and IngressRoute. Reference an existing directory (e.g. apps/fc-messageboard/) for the canonical shape.

Conventions:

• Namespace has label app.kubernetes.io/part-of: bluejay-infra
• Deployment.spec.selector.matchLabels and Service.spec.selector MUST use the same label key. The historical convention here is app: <name> (not app.kubernetes.io/name) — don't mix.
• Image: localhost/<name>:v<YYYYMMDD><HHMM>, imagePullPolicy: Never. Import the image to every RKE2 node (server + both agents) via ctr images import before applying — pods schedule anywhere.
• If the app persists local state (SQLite, uploads), declare the PersistentVolumeClaim here with storageClassName: longhorn and accessModes: [ReadWriteOnce]. Add strategy.type: Recreate to the Deployment — RWO PVC blocks rolling updates.
• Probes: use tcpSocket if the app has middleware that intercepts unauth requests (returns 404/401 for /health). Otherwise prefer httpGet against whatever the app exposes (verify the path isn't gated by auth).
• Certificate: issuerRef.name: step-ca-acme, issuerRef.kind: ClusterIssuer. dnsNames must match the hostname you created in FlowerCore.DNS in step 1.

3. Commit & push

git add apps/<name>/
git commit -m "<name>: initial deployment"
git push

ArgoCD's ApplicationSet picks up the new directory within ~3 minutes and creates infra-<name> with auto-sync + self-heal enabled.

4. Verify

# From noc1
fcadmin_ssh noc1 '
  kubectl -n argocd get application infra-<name>
  kubectl -n <ns> get certificate,pod
  curl -sk -m 8 -o /dev/null -w "HTTP %{http_code}\n" https://<name>.iamworkin.lan/
'

Certificate should be Ready: True within ~60s. If it stalls False for >2m, the pfSense DNS step got skipped — go back to step 1, then kubectl -n <ns> delete order <order-name> to bust the backoff.

Pre-merge gate

Before git push, always run:

python scripts/check-pfsense-dns.py

It's a quick service-backed check that would have caught the entire 2026-04-22 cert-manager outage. Consider wiring it into a pre-commit hook or a Gitea Actions workflow.

Retiring a service

1. kubectl -n argocd delete application infra-<name> (cascade deletes the K8s resources via ArgoCD finalizers)
2. git rm -r apps/<name>/ and push
3. Remove the FlowerCore.DNS record through the UI or API, for example:

curl -sk https://dns.iamworkin.lan/api/v1/servers
curl -sk -X DELETE https://dns.iamworkin.lan/api/v1/servers/<serverId>/zones/iamworkin.lan/records/<yourservice>

Known gotchas

• CoreDNS template + ndots:5 collision: inside pods, <svc>.<ns>.svc.cluster.local with <5 dots gets search-expanded through iamworkin.lan FIRST and hits the wildcard template → resolves to Traefik VIP, not the real ClusterIP. Use short service names (<svc>) in K8s manifests. See memory feedback_coredns_ndots_template_collision.md.
• Image not on node: pods stuck ErrImageNeverPull means the image wasn't imported to the node Kubernetes scheduled the pod onto. ctr images import on all of rke2-server, rke2-agent1, rke2-agent2.
• StatefulSet PVC drift: volumeClaimTemplates needs explicit volumeMode: Filesystem or ArgoCD SSA self-heals forever. See memory feedback_argocd_statefulset_pvc_drift.md.
• IngressRoute namespace split: this RKE2 Traefik install does not allow cross-namespace service refs. Keep the IngressRoute, backend Service, and TLS secret in the same namespace; if one host is shared across namespaces, duplicate the Certificate and move the route next to the destination service.
• Public read-only hosts: if a public host fronts a service that also exposes admin writes internally, add a Traefik route match like Host(...) && (Method(GET) || Method(HEAD)) on the public edge instead of trusting the app to reject unsafe methods.
• Public read-write allowlist hosts: if a public host accepts a tightly bounded write surface (e.g. bootstrap-JWT POST), pin the allowlist as (Method(GET) || Method(HEAD) || Method(POST) || Method(OPTIONS)). PUT/PATCH/DELETE must still 404 at the route. Track A's updatecenter.iamworkin.lan / updates.iamworkin.lan are the canonical example. The lint test enforces this invariant.
• Traefik VIP netpols: when a NetworkPolicy allows 10.0.56.200, also allow the post-DNAT backend ports (8443 for TLS plus 8080 or 8000 for HTTP) or Calico will drop the rewritten flow.
• Auth-safe probes: services behind API-key or global auth middleware should prefer tcpSocket probes unless /health is explicitly exempted before the middleware runs.
• ArgoCD must use internal Gitea URL: http://gitea-clusterip.gitea.svc.cluster.local:3000/bluejay/bluejay-infra.git, not the external HTTPS URL (step-ca cert isn't trusted by ArgoCD). The ApplicationSet and any hand-created Application must both use the internal URL.

Local manifest lint

The repo now carries a local-first lint pass for the recurring K8s gotchas that have burned the fleet:

dotnet test tests/bluejay-infra-lint/BluejayInfraLint.Tests.csproj -c Release

That test project sweeps bluejay-infra/apps/** plus the canonical sibling FlowerCore.*\\k8s manifests that share the same workspace. Matching conftest.dev policy files live under tests/bluejay-infra-lint/conftest.dev/ for environments that also have conftest or opa.

References

• OpenVox noc1 durability runbook: docs/runbooks/openvoxserver-quadlet-durability.md
• Cert-manager recovery playbook: FlowerCore.Notes/memory/project_cert_manager_recovery_2026_04_22.md
• Why pfSense DNS is required: FlowerCore.Notes/memory/feedback_pfsense_dns_required_for_acme.md
• Public DNS operator host: https://dns.iamworkin.lan
• Canonical credential helper: FlowerCore.Notes/scripts/credential-helper.sh
• pfSense admin automation: FlowerCore.Notes/memory/feedback_pfsense_automation.md