Skip to content

Appendix G — Servers in Hetzner

The Hetzner installs are in development; this is what they will need. What to rent and set up in Hetzner before running the install, and why. All of it is done by you, in Hetzner's Robot (dedicated servers) and Cloud (cloud servers) consoles: TAPPaaS never holds your Hetzner login.

How a TAPPaaS site sits in Hetzner

internet ── Hetzner's switch: accepts the server's own MAC, and the virtual MACs Robot generates
               │
            server NIC ── node: the server's main address, closed to everything from outside
                    └──── firewall VM: a second public address of its own (its virtual MAC)
                            │
                  the site's networks, inside the server (and, in a cluster, across the vSwitch)
  • The node keeps the server's main address. It accepts only ssh on it, with a key; Hetzner's rescue system stays the way in when all else fails.
  • The firewall has its own public address: an additional IP on one server, or an address from a public subnet on the vSwitch in a cluster. You, your VPN and the internet reach the site there.
  • Hetzner's switch drops, and reports as abuse, traffic from any MAC it did not hand out. That is why the firewall's address comes with a MAC generated by Robot, and why TAPPaaS never sends a made-up one.

1. The account and your SSH key

  1. A Hetzner account, with Robot and Cloud access.
  2. Robot: Key management → New key: the public half of the SSH key on your Mac or PC (~/.ssh/id_ed25519.pub). The rescue system and the install use it.

2. The server

  • Robot → Order (or the Server Auction, cheaper and immediate): x86-64, at least 32 GB of memory and NVMe disks of 0.5 TB or more. Two equal NVMe disks, the usual configuration, become a mirrored boot pool. For production: ECC memory and sizes from Appendix B.
  • Location: for a cluster, every server in the same location (a vSwitch spans one network zone, e.g. Falkenstein, Nuremberg and Helsinki form eu-central).
  • Hetzner's cloud servers cannot run TAPPaaS as a node: they do not let a server run virtual machines of its own.
  • Leave Robot's own firewall for the server off: the TAPPaaS firewall does that job.

3. The firewall's address, on a single server

  1. Robot: the server → IPs → Order additional IPs / nets: one additional IPv4. Hetzner asks what it is for (the firewall of a private cloud is a valid reason).
  2. When it is assigned, click the MAC symbol beside it to generate its virtual MAC.
  3. Note the address and the MAC: the install asks for both.

The server's IPv6 /64 is routed to its main address; the site starts with IPv4 only.

4. The rescue system, before every install

  1. Robot: the server → Rescue: Linux, 64 bit, your key → Activate rescue system.
  2. Reset → Execute an automatic hardware reset.
  3. After a minute or two, ssh root@<server address> reaches the rescue system: a Linux running in memory, with the disks untouched.

The install script does the rest from there: it builds and runs the Proxmox installer on the rescue system, writing over the server's disks, and reboots into Proxmox. The rescue system is active for one boot only.

5. For a cluster: the vSwitch and its public subnet

  1. Robot: Servers → vSwitches → Create vSwitch, with a VLAN ID from 4000 to 4091.
  2. Add every server of the site to it. Each server reaches the others over it on that VLAN (MTU 1400). TAPPaaS carries the site's own networks across it, tunnelled.
  3. In the vSwitch: IPs → order a public IPv4 subnet, /29 or larger (the network, the gateway and the broadcast take three of its addresses). It is the firewall's public address, and works on whichever server the firewall runs on.
  4. A site that started on a single server with an additional IP moves the firewall to this subnet once: its public address changes, and your DNS records and VPN endpoint with it. Cancel the additional IP afterwards.

6. For a two server cluster: the tie-breaker

Two servers need a third vote, or the survivor of a failure stops too. A small cloud server gives it:

  1. Cloud: Networks → Create network, with a range outside 10.0.0.0/8 (e.g. 172.31.0.0/16), and in it a subnet of type vSwitch, coupled to the site's vSwitch.
  2. Cloud: Servers → Add server: Debian 12 or 13, the smallest type, in the same network zone and in that network, with your SSH key.
  3. TAPPaaS makes it the cluster's QDevice (qdevice.sh add <its address>). A satellite can be the QDevice too. With three or more servers it is not needed.

7. What it costs

Hetzner's prices change; check Robot and Cloud. What a site rents:

Single server Two server cluster
Dedicated servers 1 2
Firewall address 1 additional IPv4 a public subnet on the vSwitch (/29)
vSwitch — 1 (free)
Cloud server — 1, the smallest (the tie-breaker)

Traffic on a dedicated server's uplink is included; traffic on the vSwitch counts towards it.