Step 1: Plan the cluster and prepare Debian hosts

This step should take approximately 30 to 40 minutes.

Overview

You will turn three generic Debian installations into named, reachable cluster candidates. A clear inventory is a safety control: it prevents an SFTP backup target or storage workload from silently sharing the failure domain it is meant to protect.

Before you start

Install Debian 13 on all three machines with a non-root administrator account and SSH enabled.
Assign fixed LAN addresses, preferably as DHCP reservations.
Do not install Docker, a desktop environment, a database, or a GPU driver yet.

Practice the skill

Record the physical and network inventory in the Git repository. Create docs/inventory.md; do not commit passwords, private keys, or Cloudflare tokens.

Node	Address	Role	Durable disk	Special hardware
`lab-01`	`192.168.50.21`	control plane, storage	mirrored SSD/NVMe preferred	—
`lab-02`	`192.168.50.22`	control plane, general work	SSD	—
`lab-03`	`192.168.50.23`	control plane, GPU work	SSD/NVMe model cache	NVIDIA GPU

Keep the SFTP server in a different failure domain: another host and disk, ideally with separate power or an off-site location.

Set each hostname and verify that its address does not change.
```
sudo hostnamectl set-hostname lab-01
hostnamectl
ip -brief address
```
Repeat with lab-02 and lab-03. Add all three names and addresses to your workstation's DNS or /etc/hosts until local DNS is ready.
Update the operating system and install the small set of node tools used throughout the series.
```
sudo apt update
sudo apt full-upgrade
sudo apt install --yes curl ca-certificates chrony jq nfs-common open-iscsi
sudo reboot
```
K3s and its storage add-ons require a working Linux host; installing these common dependencies now makes later failures more obvious. Reconnect after the reboot and verify timedatectl status reports synchronized time.
Decide which disk backs durable single-instance services. On lab-01, inspect disks before creating or mounting anything:
```
lsblk --fs
findmnt /
```
This tutorial uses K3s local-path storage pinned to lab-01. A disk loss is therefore a restore event, not a transparent failover. Prefer a mirrored ZFS or RAID-backed filesystem for this node, but do not confuse local mirroring with a backup.
Reserve an unused LAN range for MetalLB. Confirm it is outside your DHCP pool and is not assigned to physical devices. Record it as 192.168.50.240-192.168.50.250 in the inventory.

What you learned

A three-node control plane and a single-instance data service have different availability properties.
Backup location, power, and disk placement matter as much as backup software.
A stable address plan is prerequisite infrastructure, not documentation overhead.

Continue learning

Continue to Step 2: Harden access and establish node baselines.