homelab/project-usage/infra-practice.md

# Infrastructure Practice Playbook

Standardized procedures for troubleshooting, developing, deploying, and operating the homelab platform. Each procedure explicitly calls out where the `core` CLI fits vs `kubectl`/`helmfile`/direct cluster access. See `core-cli-tools.md` for the auth/secrets domain split; see `infra-troubleshooting.md` for quick patterns and gotchas.

## Procedure A: Troubleshoot a Service or Cluster Issue

1. **Identify the domain.**
   - Vault/secrets: `core get`/`core put` failing, Vault unreachable.
   - Node/Talos: node crashes, disk full, kubelet unreachable, network issues.
   - Plain Kubernetes: pod CrashLoop, service 503, deployment stuck.

2. **If Vault-adjacent (secrets, authentication failing):**
   - Run `core secrets status` first (NOT `core auth status` — common mistake).
   - If `Vault UNREACHABLE`, check DNS/networking:
     - No wildcard DNS exists; verify manual `/etc/hosts` entries (10.6.0.1 for WireGuard, 192.168.1.160 for LAN).
     - Ping the Vault service: `kubectl get svc -n vault | grep vault`.
   - If `Vault token not cached`, run `core secrets login`, approve device code in browser.
   - Verify: `core secrets status` shows `✓ Authenticated`.

3. **If node/Talos-adjacent (kubelet logs, node state, services failing):**
   - Run `core auth status` first.
   - If token expired, run `core auth login-oob`, approve device code in browser.
   - Then run `core nodes` to list cluster nodes.
   - For a specific node, run `core status <ip>` (Talos state).
   - Inspect Talos services: `core services <ip>` (kubelet, etcd, controller, etc.).
   - Check service logs: `core logs <ip>` (main Talos logs) or `core log-svc <ip> kubelet` (specific service).
   - Consult `infra-troubleshooting.md` for Pod stuck in CrashLoopBackOff and kubelet restart patterns.

4. **If plain Kubernetes (pod/deployment/service issues):**
   - Consult root `TROUBLESHOOTING.md` for the layer-before-tool SRE methodology (procedures 1–10).
   - Use `infra-troubleshooting.md` § Quick Patterns for common diagnoses:
     - CrashLoopBackOff: `kubectl logs -n <ns> <pod> --tail=50` + `kubectl describe pod -n <ns> <pod> | grep -A 10 Events`.
     - Service 503: `kubectl get endpoints -n <ns> <svc>` (endpoints missing?) + `kubectl get pods -n <ns> -o wide` (pods not Ready?).
     - Helm release stuck: `helmfile status | grep -E "FAILED|UNKNOWN|PENDING"` + `helm status <release> -n <ns> --show-resources`.
   - If still unclear, escalate to `kubectl get all -n <ns>` and review resource events.

5. **For dashboards and live metrics:**
   - Use `core pf grafana` (port-forward to localhost:3000) rather than raw `kubectl port-forward` — keeps forwarded ports consistent.
   - If Prometheus unavailable, check: `kubectl get pods -n monitoring | grep prometheus`.
   - If ServiceMonitor not scraping, verify: `kubectl get servicemonitor -A | grep <name>` and inspect `.spec.selector` matches the target pod's app label.

---

## Procedure B: Launch/Develop a New Service or POC

1. **Plan the service.**
   - Determine namespace (e.g., `sqs`, `temporal`, `databases`, `monitoring`).
   - Decide if metrics exported (most should) and if OIDC-gated.
   - Sketch a Helm values.yaml structure (secrets, replicas, resource requests, affinity).

2. **Authenticate to Vault.**
   - Run `core secrets login` and approve device code in browser.
   - Verify: `core secrets status` shows `✓ Authenticated`.
   - You'll need Vault access to store service secrets in step 5.

3. **Create service Helm chart directory.**
   - Create `k8s/<service>/` with at minimum:
     - `values.yaml` (Helm values for deployment, service, replicas, resource limits).
     - `charts/` subdirectory for any custom local Helm charts (optional).
   - Follow naming conventions from `coding-standards.md`.

4. **Add Helm release to helmfile.**
   - Open `helmfile.yaml.gotmpl`.
   - Add release block under `releases:` section, following this structure:
     ```yaml
     - name: <service>
       namespace: <namespace>
       chart: <chart-repo>/<chart-name>
       version: ~1.0  # pin major.minor, allow patch updates
       needs:
         - <dependency-namespace>/<dependency-release>  # if applicable
       values:
         - k8s/<service>/values.yaml
         - secretsInline:
             DB_PASSWORD: "{{ env \"<SERVICE>_DB_PASSWORD\" }}"
     ```
   - Consult `coding-standards.md` for `needs:` ordering (example: sqs section shows strimzi-operator → kafka-cluster → queue-crd → management-service).
   - Reference real example: root helmfile's `sqs` section.

5. **If the service needs secrets (DB password, API key, OAuth secret):**
   - Generate value (e.g., `openssl rand -hex 32` for passwords).
   - Store in Vault: `core put cluster/<SERVICE>_<KEY> <SERVICE>_<KEY>="value"`.
   - **Critical gotcha:** field name MUST equal variable name (e.g., `FORGEJO_ADMIN_PASSWORD=` not `value=`) per `coding-standards.md` § Vault field=variable convention.
   - Reference in values.yaml via `{{ env "VARIABLE_NAME" }}` (Helmfile Go template syntax, NOT shell `${VAR}`).
   - Do NOT hardcode secrets in values.yaml or ConfigMaps.

6. **Verify Helm syntax before deploy.**
   - Run `helmfile lint` (catches template errors, duplicate releases).
   - Run `helmfile diff -l name=<service>` (show what will be deployed).
   - Review diff for correctness (verify env var substitutions, resource limits, affinity rules).

7. **Deploy the service.**
   - Run `helmfile apply -l name=<service>`.
   - Monitor: `kubectl get pods -n <namespace> -w` (watch until Running).
   - If pods stuck: `kubectl describe pod -n <namespace> <pod-name>` (check Events for SchedulingFailed, ImagePullBackOff, etc.).

8. **If the service exports `/metrics` (Prometheus format):**
   - Create ServiceMonitor: `k8s/monitoring/servicemonitors/svc-<name>.yaml`.
     - `.spec.selector.matchLabels` must match the service's pod labels (usually `app: <service>`).
     - `.spec.endpoints[0].port` must match the service port name or number exporting metrics.
   - Create PrometheusRule: `k8s/monitoring/alerts/svc-<name>-rules.yaml`.
     - Include error rate, latency, and SLO alert rules.
     - Use `prometheus` as the rule group.
   - Create Grafana dashboard: `k8s/monitoring/dashboards/svc-<name>.yaml`.
     - Use 6-row template: Availability, Resources, Domain metrics, Logs, SLO, Related.
     - See README.md § Example Applications for a full walkthrough.
   - Verify scrape: `kubectl get servicemonitor -A | grep <name>` and check Prometheus Targets UI for green status.

9. **If OIDC/IAM-gated (admin UI, restricted API):**
   - Create app in Authentik: `core iam create-app "my-service" --slug my-service --redirect-uri "https://my-service.riotpiao.com/callback"`.
   - Bind app to group: `core iam bind-app my-service <group>` (e.g., `grafana-admins` for admin-only UI).
   - Retrieve credentials: `core iam describe-app my-service` (client ID, client secret).
   - Deploy secret: `kubectl create secret generic <service>-oidc --from-literal=client-id=<ID> --from-literal=client-secret=<SECRET> -n <namespace>`.
   - Reference secret in values.yaml: mount via `.spec.template.spec.containers[].env` or volumeMounts.
   - See `core-cli-tools.md` § Access Control Tiers for Tier A (OIDC + RBAC) vs Tier B (network perimeter only).

10. **Verify service is live.**
    - Pods: `kubectl get pods -n <namespace> -o wide` (all Running, 1/1 Ready).
    - Metrics (if applicable): `kubectl get servicemonitor -A | grep <name>` and visit Prometheus Targets or Grafana dashboard.
    - Endpoint: If publicly routed via Ingress, verify `/etc/hosts` entry (10.6.0.1 for WireGuard, 192.168.1.160 for LAN) and `curl https://my-service.riotpiao.com/health` (or equivalent health endpoint).
    - Logs: `kubectl logs -n <namespace> <pod>` (no errors).

### Definition of Done (Per Service)

- [ ] Helm chart version pinned (~1.0 format in helmfile)
- [ ] All secrets in Vault (none in values.yaml or ConfigMap)
- [ ] `/metrics` endpoint exported (if applicable)
- [ ] ServiceMonitor resource created (if metrics exported)
- [ ] PrometheusRule with error/latency/SLO alerts (if metrics exported)
- [ ] Grafana dashboard (if metrics exported; 6-row template: Availability, Resources, Domain, Logs, SLO, Related)
- [ ] Ingress rule (if external access needed)
- [ ] OIDC integration via `core iam` (if UI component)
- [ ] Verified: `helmfile diff` clean, pods Running, dashboard live or `/metrics` returning 200

---

## Procedure C: Operate the Cluster (Node Health, Context, Cleanup)

1. **Daily health check.**
   - Check auth: `core auth status` (if OK, node ops will work).
   - List nodes: `core nodes`.
   - For each node, check Talos state: `core status <ip>`.
   - Check K8s nodes: `kubectl get nodes -o wide` (all Ready, no NotReady).
   - Check pod pressure: `kubectl get nodes -o json | jq '.items[] | {name: .metadata.name, memory: .status.allocatable.memory, pods: .status.allocatable.pods}'`.

2. **Troubleshoot a specific node.**
   - Get node IP: `core nodes` and note the IP.
   - Check Talos services: `core services <ip>` (kubelet, etcd, controller should be running).
   - Check service logs: `core logs <ip>` (main Talos daemon logs).
   - Filter to specific service: `core log-svc <ip> kubelet` (kubelet logs only).
   - Restart a service if needed: `core restart <ip> kubelet` (graceful kubelet restart).

3. **Pod cleanup (Failed, Evicted, Terminating pods).**
   - Run `core pods clean` (scans all namespaces, removes stale pods).
   - Verify: `kubectl get pods -A | grep -E "Failed|Evicted"` (should be empty).

4. **Switch kubectl context (when off-LAN, on WireGuard).**
   - List available contexts: `core config kube-list`.
   - Switch to WireGuard path (10.6.0.1:6443): `core config kube-use admin@homelab-cluster-1`.
   - **Known limitation:** `core config use <talos-context>` doesn't map to WireGuard; use `kube-use` directly.
   - Verify: `kubectl cluster-info` shows 10.6.0.1 (not 192.168.1.213).

5. **MinIO bucket operations (if managing data/backups).**
   - List buckets: `core bucket list`.
   - Upload file: `core bucket upload <bucket> <local-file>`.
   - Download file: `core bucket download <bucket> <remote-file> -o <local-file>`.
   - Delete file: `core bucket delete <bucket> <remote-file>`.

6. **Bootstrap or hardware runbooks (infrequent).**
   - **Fresh cluster setup:** See README.md § Bootstrap Order (14 steps).
   - **Adding a new Talos node:** See README.md § Adding Hardware.
   - Do not re-explain those long procedures here; consult README.md directly.

---

## Notes

**Queue subsystem (Kafka/kmsvc/Temporal namespace auto-registration):** Already deployed and stable. If re-deploying:
- Primary deploy method: `helmfile apply -l namespace=sqs` (live from root helmfile).
- Alternate isolated iterate path: `k8s/sqs/helmfile.yaml.gotmpl` (not recommended for production).
- Planned future: GitOps via `k8s/sqs/argocd/` (companion repo, not yet active).
- **Critical rule:** Temporal namespace registration is automatic via `queue-operator`; never manually `temporal operator namespace create` for any namespace referenced by a Queue's `temporal.io/namespace` label. See `~/workplace/kmsvc-manage/CLAUDE.md` ("Temporal Namespace Registration") for the full rule and why.

**Shared/reusable service repositories:** If a service's Helm chart and container image live in a separate repository, they must be:
- Published as a public GitHub repository under the `Riotpiaole` organization.
- Images pushed to GHCR (`ghcr.io/riotpiaole/...`) for public pullability.
- Consult `coding-standards.md` § Shared/Reusable Repos for the full publishing rule.

---

## Cross-References

- **core-cli-tools.md:** Auth/secrets domain split, command inventory, when to use `core` vs `kubectl`.
- **coding-standards.md:** Helm naming conventions, `needs:` ordering rules, helmfile template syntax (`{{ env "VAR" }}` not `${VAR}`), Vault field=variable convention, shared-repo publishing rule.
- **infra-troubleshooting.md:** Quick patterns (CrashLoopBackOff, 503, helm stuck), gotchas, hard rules.
- **USAGE.md:** Exhaustive `core` command reference.
- **README.md:** Bootstrap order, hardware addition, example app walkthrough, 6-row Grafana dashboard template.
- **root TROUBLESHOOTING.md:** Generic Kubernetes SRE layer-before-tool methodology (10 diagnostic procedures).