Self-Host Fleet (No Kubernetes)

You've outgrown one VPS but Kubernetes is overkill. This is the sweet spot a huge number of profitable companies actually live in.

Components

• Hardware: 3–10 VPSes from Hetzner / Hetzner Cloud / OVH / Vultr / DigitalOcean. Or bare metal at Hetzner Robot ($30–80/mo for huge boxes).
• Orchestration: Kamal 2 (Basecamp's tool — Docker-based deploys to any SSH-able box). Or Ansible for fuller config management.
• Reverse proxy: Caddy on each app server, or HAProxy / Nginx if you need fancier load balancing.
• DB: Managed Postgres (Neon / Crunchy / RDS) or self-hosted with Patroni for HA.
• Cache / queue: Redis / DragonflyDB on a dedicated box, or use Postgres for both (pg-boss).
• Object storage: Cloudflare R2 (no egress) or Backblaze B2.
• Network: Tailscale mesh; only 80/443 publicly exposed.
• Observability: Prometheus + Grafana + Loki on its own VPS, agents on every host.
• Backup: restic → R2 + DB-level WAL archiving.
• Secrets: SOPS + age, or Doppler / Infisical (see web-dev secrets).
• CI/CD: GitHub Actions → Kamal deploy / Ansible playbook.

What it's good at

• Cost-effective at scale. A few Hetzner bare-metal boxes cost less than one similarly-sized AWS instance.
• Predictable performance. Real CPUs, real disks, no noisy neighbors.
• Total control. Custom kernels, weird storage, on-disk encryption — all yours.
• Clear failure modes. When something breaks you can ssh in and look.

Where it strains

• Ops time. Patching, log rotation, hardware failure, network blips — they're your problem.
• HA storage. Real Postgres HA with Patroni is doable but non-trivial.
• Compliance. SOC 2 evidence collection is harder without managed services.
• Burst capacity. Hard to add and remove servers in minutes.

Tooling that makes this stack pleasant

• Kamal 2 — one config file deploys Docker images to any SSH host with zero-downtime rollouts.
• Ansible — for actual host config (users, packages, sysctl, kernel tuning).
• Tailscale — collapse the "VPN + bastion + jump-host" complexity into one daemon.
• Prometheus + Grafana + Loki — monitoring without paying Datadog.
• Hetzner snapshot API + Terraform — recreate the whole fleet in 20 minutes.

Common topology

• 1 reverse-proxy / load balancer box (Caddy or HAProxy).
• 2–N app boxes (Kamal-deployed, identical).
• 1 DB box (or managed Postgres).
• 1 background-jobs box.
• 1 monitoring / log box.

Multiply or split as needed. With a load balancer + 2 app boxes you can patch one while the other serves.

When to evolve

• Cluster grows past ~15 boxes → consider Nomad or Kubernetes.
• Multiple teams deploying to overlapping infra → IDP / k8s.
• Need true autoscaling → managed cloud or k8s.

Pick this if…

• 3–20 services, 2–10 engineers, modest growth.
• You have at least one person who enjoys ops.
• You'd like your AWS bill to look like a phone bill, not a mortgage.
• Your data + compliance posture benefits from full control.

Skip if your team isn't comfortable in a terminal, or if your compliance team would prefer "the cloud handles that."