DEPLOYMENT

Production Reference Architecture

Scale, upgrades, recovery, and controller failure handling.

Reference stack

Tier	Components
Edge	Ingress (nginx/traefik), TLS, WAF optional
Control	3× aether serve pods, anti-affinity
Data	PostgreSQL (state), Redis (sessions)
Targets	EKS/GKE/AKS/on-prem via kubeconfig
Observability	Prometheus scrape /metrics, structured JSON logs

Install via helm install aether ./helm/aether with HA values.

---

Sizing (starting point)

Workloads	API replicas	Postgres
< 100	2	2 vCPU / 4Gi
100–500	3	4 vCPU / 8Gi
500+	3+ horizontal	Managed RDS + read replica

API rate limit: 200 concurrent requests (Tower middleware).

---

Upgrades

1. Backup state: aether backup or Postgres snapshot
2. Rolling update Helm release
3. Verify /api/health and dashboard SSE
4. Run smoke: aether list, sample migrate dry-run

---

Controller failure recovery

Failure	Recovery
API pod crash	Kubernetes restarts pod; state in Postgres
Postgres loss	Restore from backup; replay GitOps
Redis loss	Sessions cleared; users re-login
Full cluster loss	Restore Postgres + redeploy Helm; workloads re-run from specs

Workload runtime state lives on target clusters — Aether control plane loss does not delete Pods/VMs.

---

Backup cadence

• State DB: daily automated snapshot
• ~/.aether (dev): copy off-host weekly
• GitOps repo: source of truth for specs

See RUNBOOK.md, BACKUP.md.

---

Observability

Tier	Components
Metrics	Prometheus scrape /api/metrics; set AETHER_PROMETHEUS_URL on API pods
Dashboards	Grafana + scripts/import-grafana-dashboard.sh; set AETHER_GRAFANA_URL and AETHER_GRAFANA_DASHBOARD_UID
Discovery	eval "$(./scripts/wire-observability-env.sh)" after installing a monitoring stack
Cilium	Bootstrap policies via deploy; connectivity ConfigMap aether-cilium-connectivity

On k3s/kind deploys, AETHER_INSTALL_METRICS_SERVER=auto (default) installs metrics-server when missing. See CILIUM.md.

---

Related

• Deployment Topologies
• AIR-GAPPED.md