# Server

Infrastructure-as-code and service definitions for luke-else.co.uk's self-hosted server estate: three [Hetzner Cloud](https://www.hetzner.com/cloud/) VPS instances provisioned with [OpenTofu](https://opentofu.org/), each running a set of Docker Compose stacks behind [Traefik](https://traefik.io/traefik/).

<p align="center">
  <img src="assets/images/main.png" width="70%">
</p>

## Contents

- [Architecture](#architecture)
- [Repository layout](#repository-layout)
- [Prerequisites](#prerequisites)
- [Provisioning the infrastructure](#provisioning-the-infrastructure-infra)
- [Deploying the services](#deploying-the-services-services)
- [Service inventory](#service-inventory)
- [First-time setup](#first-time-setup)
- [Development container](#development-container)
- [Security notes](#security-notes)

## Architecture

Three servers, one shared private network:

```mermaid
architecture-beta
    group cloud(cloud)[Hetzner]

    group network(cloud)[network] in cloud

    service disk1(mdi:disk)[Storage] in cloud
    service disk2(mdi:disk)[Storage] in cloud

    service dev(mdi:server)[dev] in network

    service prod(mdi:server)[prod] in network
    service prodfirewall(mdi:firewall)[firewall] in cloud

    service vpn(mdi:server)[vpn] in cloud
    service vpnfirewall(mdi:firewall)[firewall] in cloud

    service gateway(mdi:web)[gateway] in cloud

    dev:L -- R:prod
    disk1:B -- T:prod
    disk2:B -- T:dev

    prod:B -- T:prodfirewall
    vpn:B -- T:vpnfirewall

    prodfirewall: L -- R: gateway
    vpnfirewall: B -- T: gateway
```

`gateway` represents the public internet, not a provisioned resource.

| Server | Purpose | Network | Volume |
|---|---|---|---|
| `dev`  | Gitea, CI runner, dev-facing Traefik | Private network only (no public firewall exposure beyond CI/CD) | `dev-storage` |
| `prod` | Public-facing websites, Bitwarden, RustDesk, status page, prod Traefik | Private network + public firewall | `prod-storage` |
| `vpn`  | OpenVPN + its own Traefik | **Not** attached to the private network — kept isolated so a compromised VPN endpoint can't pivot to `dev`/`prod` | none |

`dev` and `prod` share a private Hetzner network (`10.0.1.0/24` by default) so CI/CD on `dev` can reach deployment targets on `prod` without exposing that traffic publicly. `vpn` is deliberately kept off this network. Each server has its own Hetzner Cloud Firewall (see `infra/modules/<host>/main.tf`) that only opens the ports actually used by the compose stacks running on it, plus SSH restricted to `var.allowed_ssh_source_ips`.

## Repository layout

```
.
├── infra/                    # OpenTofu (Terraform-compatible) config — provisions the 3 servers, network, firewalls, volumes
│   ├── main.tf                # root module: wires network + dev/prod/vpn modules together
│   ├── variables.tf           # shared inputs (sizes, locations, IP ranges, SSH key names)
│   ├── outputs.tf             # pass-through outputs from each host module
│   ├── ssh.tf                 # looks up each SSH key already uploaded to Hetzner Cloud
│   ├── versions.tf            # provider requirements
│   ├── terraform.tfvars.example
│   ├── scripts/
│   │   └── bootstrap.sh.tftpl # post-install script run on every host right after creation
│   └── modules/
│       ├── network/           # shared private network + subnet (used by dev and prod)
│       ├── dev/                # dev server + firewall + volume, renders + deploys services/dev/
│       │   └── templates/
│       │       └── runners-docker-compose.yml.tftpl # shape of each generated runner service
│       ├── prod/               # prod server + firewall + volume, deploys services/prod/
│       ├── vpn/                # vpn server + firewall (no private network, no volume), deploys services/vpn/
│       └── dns/                # Hetzner DNS zones + A records for every domain in var.dns_zones
├── services/                 # Docker Compose stacks, grouped by which server they run on
│   ├── dev/                   # Gitea + CI runner + Traefik (git.luke-else.co.uk, cicd.luke-else.co.uk)
│   ├── prod/                  # Websites, Bitwarden, RustDesk, status page + Traefik
│   ├── vpn/                    # OpenVPN + Traefik
│   └── todo.md                # Outstanding manual setup/hardening tasks
├── docs/
│   └── architecture.md       # Source of the architecture diagram above
├── .devcontainer/             # Git submodule: shared devcontainer for working on this repo (OpenTofu tooling)
└── assets/
```

Each of `services/dev`, `services/prod`, `services/vpn` follows the same convention: one `*-docker-compose.yml` (or subdirectory, e.g. `Runners/`) per logical service, plus a `spinup.sh` / `spindown.sh` pair that brings up or tears down every stack on that host in the right order. `infra/modules/` mirrors this same dev/prod/vpn split on the provisioning side, so a given host's cloud resources and its compose stacks are easy to find side by side.

## Prerequisites

- A [Hetzner Cloud](https://console.hetzner.cloud/) project and API token
- One or more SSH keys uploaded to that project (Console → Security → SSH Keys) - all are installed on every server - plus the private key matching one of them available locally (OpenTofu uses it once per server to run the post-install bootstrap — see below)
- [OpenTofu](https://opentofu.org/docs/intro/install/) `>= 1.6.0`
- Ownership of the domains in `var.dns_zones` at whatever registrar they're bought through, so you can point their NS records at Hetzner (see [Managing DNS](#managing-dns) — the zones and records themselves are created for you)

Docker + the Compose plugin no longer need installing by hand — the bootstrap script below handles that.

## Provisioning the infrastructure (`infra/`)

```sh
cd infra
export HCLOUD_TOKEN=your-hetzner-api-token   # never commit this
cp terraform.tfvars.example terraform.tfvars
$EDITOR terraform.tfvars                     # set ssh_key_names and ssh_private_key_path at minimum

tofu init
tofu plan
tofu apply
```

This creates, via `module.network` / `module.dev` / `module.prod` / `module.vpn` / `module.dns` in `infra/main.tf`:
- `hcloud_network` + subnet, shared by `dev` and `prod` (`modules/network`)
- one `hcloud_server` + `hcloud_firewall` per host, scoped to the ports each host actually uses (`modules/dev`, `modules/prod`, `modules/vpn`)
- one `hcloud_volume` each for `dev` and `prod` (`modules/dev`, `modules/prod`; `vpn` has none)
- one Hetzner DNS zone per domain in `var.dns_zones`, plus every A record in [Service inventory](#service-inventory) (`modules/dns` — see [Managing DNS](#managing-dns))

Useful outputs: `tofu output dev_ipv4`, `tofu output prod_ipv4`, `tofu output vpn_ipv4`, `tofu output dns_nameservers`, `tofu output dev_data_dir`, `tofu output prod_data_dir`.

`terraform.tfvars` and any `*.tfvars` file are gitignored — never commit real values there. Defaults for server sizes, locations, and IP ranges live in `infra/variables.tf` and are passed down into the modules from `infra/main.tf`; override them per-environment via `terraform.tfvars`.

### Post-install bootstrap

Immediately after each `hcloud_server` is created, OpenTofu uploads and runs [`infra/scripts/bootstrap.sh.tftpl`](infra/scripts/bootstrap.sh.tftpl) over SSH as `root` (`connection` + `file`/`remote-exec` provisioners in each `modules/<host>/main.tf`). It:

- installs Docker Engine + the Compose plugin
- creates a non-root sudo user (`var.deploy_user`, default `deploy`) with the same SSH keys as root, in the `sudo` and `docker` groups
- hardens `sshd`: password authentication off, root login restricted to key-only (`PermitRootLogin prohibit-password`)
- installs and enables `unattended-upgrades` for automatic security patches

After bootstrap, SSH in as `var.deploy_user` for day-to-day work — root key-based login still works as a fallback.

### Persistent data lives on the volumes, not the server disk

`dev` and `prod`'s compose files used to bind-mount container data to paths relative to wherever `services/<host>/` happened to be copied on the server (e.g. `./gitea:/data`) - i.e. the server's local disk, which doesn't survive the server being destroyed and recreated. `dev-storage` and `prod-storage` (the `hcloud_volume`s) do survive that, so all stateful bind mounts now point there instead: `${DATA_DIR}/gitea:/data`, `${DATA_DIR}/bitwarden/:/data/`, and so on across `services/dev/*.yml` and `services/prod/*.yml`. `vpn` has no volume, so it's untouched - see the architecture table above.

`DATA_DIR` needing to resolve to the right path is what makes this work, and that's answered deterministically: Hetzner's automount tooling always mounts an auto-mounted volume at `/mnt/HC_Volume_<volume-id>`, and since the volume is a Terraform resource, `modules/<host>/main.tf` already knows its `id` the moment it's created - `local.data_dir = "/mnt/HC_Volume_${hcloud_volume.storage.id}"`. Every `tofu apply` writes that value straight into `services/dev/.env` / `services/prod/.env` as `DATA_DIR=...`, which `docker compose` auto-loads for variable substitution in any compose file run from that directory (the same mechanism the Gitea runner token already uses). Query it directly with `tofu output dev_data_dir` / `tofu output prod_data_dir`.

These `.env` files are deliberately not committed (see `.gitignore`): the path is tied to the *live* volume's ID, so it's regenerated by `tofu apply`, not handed off via git. Run `tofu apply` at least once before the first deploy, and again if a volume is ever destroyed and recreated (new ID); `spinup.sh` on both hosts refuses to start anything if `.env` is missing, rather than silently falling back to a relative path.

The Gitea Actions runners are the one exception to the `.env` mechanism: their `/data` path is baked directly into the generated `Runners/docker-compose.yml` at render time (via the same `templatefile()` call as `dev_runner_count`), because `Runners/.env` is already reserved for the registration token and gets overwritten on every deploy.

### Scaling Gitea Actions runners

The number of Gitea Actions runner containers on `dev` is an OpenTofu input, `var.dev_runner_count` (default `3`). On every `tofu apply`, `modules/dev`'s `local_file.runners_compose` resource renders [`modules/dev/templates/runners-docker-compose.yml.tftpl`](infra/modules/dev/templates/runners-docker-compose.yml.tftpl) straight into [`services/dev/Runners/docker-compose.yml`](services/dev/Runners/docker-compose.yml) in your working tree — one `runner-N` service per count, each with its own container name and `/data` volume so their registrations don't collide. That file is generated: change `dev_runner_count` in `terraform.tfvars` and re-run `tofu apply` rather than hand-editing it.

This only updates your local working tree — `tofu apply` doesn't reach out and start containers on `dev` itself. Redeploy as usual (re-copy `services/dev/` to the server and re-run `spinup.sh`) to apply a count change.

Registration tokens are handled automatically, not baked into the generated file: `services/dev/spinup.sh` waits for Gitea to come up, runs `gitea actions generate-runner-token` inside the Gitea container, and writes the result to `Runners/.env`, which Compose loads automatically. There's no manual admin-UI step for this anymore.

### First apply: bootstrapping order matters

`dev` and `prod`'s firewalls only accept SSH from `vpn`'s public IP (see [Security notes](#security-notes)), but `vpn`'s own OpenVPN service isn't running until you deploy it — so on a from-scratch `tofu apply`, the `dev`/`prod` bootstrap provisioners can't connect yet. Bring the estate up in this order:

1. `tofu apply -target=module.vpn` — creates `vpn` only; its firewall still allows SSH from `var.allowed_ssh_source_ips`, so its bootstrap provisioner and `null_resource.deploy_services` (which copies `services/vpn` over) both run fine.
2. SSH in as `var.deploy_user` and run `~/services/vpn/spinup.sh` to start the VPN service, then connect to it with an OpenVPN client.
3. `tofu apply` — creates `network`, `dev`, `prod`. Now that your machine is tunneled through `vpn`, its NATed egress IP matches the firewall rule and the `dev`/`prod` bootstrap + deploy provisioners can connect.

If step 3 is run before you're connected to the VPN, the `dev`/`prod` provisioners will fail and Terraform will mark those servers tainted — re-running `tofu apply` once connected will recreate and re-bootstrap them.

### Managing DNS

`module.dns` (in `infra/modules/dns`) creates one Hetzner DNS zone per domain in `var.dns_zones` (default: `luke-else.co.uk`, `divine-couture.co.uk`, `snexo.co.uk`) and an `hcloud_zone_rrset` A record for every hostname currently referenced by a Traefik `Host()` rule anywhere under `services/` — kept in sync with `local.dns_records` in `infra/main.tf`, pointed at whichever of `dev`/`prod`/`vpn` actually serves it. Zones have `prevent_destroy = true`, matching the volumes — losing one deletes every record in it.

Creating the zone doesn't make Hetzner authoritative for the domain by itself: you still need to point that domain's NS records at Hetzner's nameservers at whichever registrar it's registered through. Run `tofu output dns_nameservers` after applying to get the exact nameservers per domain, and set those as the domain's NS records at the registrar. Propagation can take a while depending on the registrar and the domain's previous NS TTL.

Adding a new subdomain: add an entry to `local.dns_records` in `infra/main.tf` (zone, name, and the target `module.<host>.ipv4`) and re-run `tofu apply` — don't hand-create records in the Hetzner console, they'll drift from state. Note `name = "@"` is Hetzner's convention for a zone's apex record (e.g. bare `snexo.co.uk`), not an empty string.

Only A records for IPv4 are managed here — none of the modules currently track servers' IPv6 addresses, so AAAA records aren't generated even though the firewalls already allow IPv6 traffic.

## Deploying the services (`services/`)

`tofu apply` already puts `services/<host>/` on the matching server for you — a `null_resource.deploy_services` in each `modules/<host>/main.tf` copies it to `/home/<var.deploy_user>/services/<host>` and `chown`s it to `var.deploy_user`, re-running whenever the server is recreated or the directory's contents change (a hand-edited compose file, a new `dev_runner_count`, a refreshed `DATA_DIR`, ...) - not just once at first creation. So after applying, just SSH in as `var.deploy_user` and run the matching script from inside that directory:

```sh
# on dev
./spinup.sh     # Traefik → Gitea → (waits, generates a runner token) → Runners → Watchtower
./spindown.sh   # reverse order, then prunes images/volumes

# on prod
./spinup.sh    # Traefik, → Watchtower → status → websites → Bitwarden → RustDesk
./spindown.sh

# on vpn
./spinup.sh     # Traefik → OpenVPN → Watchtower
./spindown.sh
```

If you ever need to force a re-sync without going through Terraform (e.g. testing a local edit before committing it), a manual `scp -r services/prod deploy@<prod_ipv4>:~/services` still works fine - `tofu apply` will just overwrite it again next time triggers change.

Each stack can also be managed individually with plain Compose, e.g.:

```sh
cd services/prod
docker compose -f bitwarden-docker-compose.yml up -d
docker compose -f bitwarden-docker-compose.yml down
```

All three hosts run [Watchtower](https://containrrr.dev/watchtower/) polling every 60s with cleanup enabled, so images are kept current automatically once deployed — the spinup scripts only need to be re-run after adding/removing a service or changing compose files.

Every public-facing service is fronted by its host's own Traefik instance, terminating TLS via Let's Encrypt (`tlschallenge`, port 80/443). Each stack joins an external `proxy` Docker network and opts in via `traefik.enable=true` labels rather than publishing ports directly (RustDesk and the Gitea SSH port are the deliberate exceptions, since they aren't HTTP).

## Service inventory

### `dev`

| Service | Compose file | Domain(s) |
|---|---|---|
| Traefik | `traefik-docker-compose.yml` | `traefik.cicd.luke-else.co.uk` |
| Gitea | `gitea-docker-compose.yml` | `git.luke-else.co.uk` (HTTP), SSH on `222` |
| Gitea Actions runner(s) | `Runners/docker-compose.yml` (generated — see [Scaling Gitea Actions runners](#scaling-gitea-actions-runners)) | N/A |
| Watchtower | `watchtower-docker-compose.yml` | — |

### `prod`

| Service | Compose file | Domain(s) |
|---|---|---|
| Traefik | `traefik-docker-compose.yml` | `traefik.luke-else.co.uk` |
| Websites | `web-docker-compose.yml` | `luke-else.co.uk`, `dev.luke-else.co.uk`, `metarius.luke-else.co.uk`, `www.divine-couture.co.uk`, `snexo.co.uk` |
| Status page (Uptime Kuma) | `status-docker-compose.yml` | `status.luke-else.co.uk` |
| Bitwarden (Vaultwarden) | `bitwarden-docker-compose.yml` | `bitwarden.luke-else.co.uk` |
| RustDesk relay (hbbs/hbbr) | `rd-docker-compose.yml` | `rd.luke-else.co.uk`, ports `21115-21119` |
| Watchtower | `watchtower-docker-compose.yml` | — |

### `vpn`

| Service | Compose file | Domain(s) |
|---|---|---|
| Traefik | `traefik-docker-compose.yml` | `traefik.vpn.luke-else.co.uk` |
| OpenVPN (Dockovpn) | `vpn-docker-compose.yml` | `vpn.luke-else.co.uk`, UDP `1194` |
| Watchtower | `watchtower-docker-compose.yml` | — |

## First-time setup

A few things need manual attention before a stack is fully live — tracked in [`services/todo.md`](services/todo.md), summarized here:

- **General host hardening**: non-root user, Docker, and unattended-upgrades are now handled automatically by the [post-install bootstrap](#post-install-bootstrap); UFW is the remaining manual item in `services/todo.md` (the Hetzner Cloud Firewalls already allowlist per-host ports — see [Security notes](#security-notes)).

## Development container

`.devcontainer` is a git submodule providing a ready-to-use OpenTofu development environment (VS Code + OpenTofu/Docker/Mermaid extensions). After cloning:

```sh
git submodule update --init --recursive
```

Then reopen the repo in VS Code with the Dev Containers extension.

## Security notes

- Real secrets (`HCLOUD_TOKEN`, `*.tfvars`, your SSH private key) must never be committed — see `.gitignore`.
- SSH to `dev` and `prod` is restricted to the `vpn` server's own public IP — you must be tunneled into the VPN to reach them over SSH. SSH to `vpn` itself is gated by `var.allowed_ssh_source_ips`; narrow this from the default `0.0.0.0/0` once you know your own egress IP(s).
- `vpn` is intentionally excluded from the private network so that a compromised VPN endpoint cannot reach `dev` or `prod` directly over it — SSH access still works because the VPN's egress traffic is NATed through its own public IP.
- Firewalls are allowlists scoped per host in `infra/modules/<host>/main.tf` — only ports actually used by that host's compose stacks (plus the cross-host private network range) are open.
- Every host's [post-install bootstrap](#post-install-bootstrap) disables SSH password authentication, restricts root login to key-only, and creates a separate sudo user (`var.deploy_user`) for day-to-day access.
- DNS zones (`modules/dns`) carry both Hetzner's own `delete_protection` and Terraform's `prevent_destroy` — losing a zone takes every record in it with it, including for `divine-couture.co.uk` and `snexo.co.uk`, not just `luke-else.co.uk`.