What problem it solves
A provider creates a server in a minute, and in that same minute bots start knocking on its address: data-center ranges are scanned nonstop. The provider’s image is built so you can log in, not so it can live on the internet for years.
How the server arrives
- login as
rootwith a password from an email or a key from the panel - port 22 open to the whole internet, password login allowed
- packages as of the image build, sometimes six months old
- UTC time zone, a name like
vm-4821 - no firewall, no swap, a journal that grows to 10% of the disk (capped at 4 GB)
How we leave it
- your own
sudouser, key-only login, root closed over SSH - only 22, 80 and 443 open to the outside
- security updates install themselves every night, with a 4:00 reboot when needed
- 2 GB of swap, a journal capped at 500 MB, the clock in sync
- Docker from Docker’s repository, container logs capped at 30 MB
- an address that guesses logins gets banned across the whole server
The order
The steps depend on each other: you can’t close passwords until key login is tested, and you can’t turn the firewall on until it has a rule for SSH. Hence this order:
server from the provider panel · root · password
- Fallback access
provider consoleBefore any change, check that the web console in the panel works. It bypasses the network and SSH and saves you if you lock yourself out (Provider console). - Access
adduser · sudo · keyUpdate packages, create a user, add a key, log in with it, close passwords and root (First login to a server). - Firewall
ufw · 22, 80, 443Rules first, thenufw enable. ufw doesn’t close Docker’s ports — that’s a separate story (Docker ports bypass the input chain). - Updates
unattended-upgradesAlready on in Ubuntu, but without reboots: a new kernel waits until someone boots it (unattended-upgrades). - System
swap · journald · timeSwap for a small VPS, a journal limit, a clock check (Swap, journal and time). - Docker
docker-ce · daemon.jsonFrom Docker’s repository, with a log limit (Docker on Ubuntu). Before the first container: the limit indaemon.jsonapplies only to containers created after it. - Brute-force bans
CrowdSecCrowdSec in a container bans an address that guesses logins on every port of the server (CrowdSec against SSH brute force).
a server you can leave on the internet
Install Docker in the first hour, not next to the first site: the services you add yourself from here on —
CrowdSec, Traefik, sites and apps — run in containers. Each containerized service lives in its own directory
with a compose file. The image tag sets the version. Update a service with docker compose pull and up -d;
remove it with down and delete the directory. Service dependencies don’t mix with system packages. That’s why such a server can be rebuilt
from files (Server configs in git). Next come the proxy and the first site (Traefik).
Ubuntu 24.04 or 26.04
Both are LTS, both get 5 years of standard support plus 5 more with Ubuntu Pro (free for personal use on up to 5 machines). The differences you’ll notice when setting up a server:
| 24.04 LTS “Noble Numbat” | 26.04 LTS “Resolute Raccoon” | |
|---|---|---|
| release and support | April 2024, updates until April 2029 | April 2026, updates until April 2031 |
| kernel | 6.8 (newer via HWE) | 7.0 |
| OpenSSH | 9.6: no PerSourcePenalties |
10.2: hybrid ML-KEM key exchange by default, DSA keys rejected |
sudo |
classic sudo | sudo-rs in Rust; sudo -E doesn’t keep the environment — you need --preserve-env=VAR; the password shows as asterisks |
| time | systemd-timesyncd, plain NTP | chrony with NTS (encrypted NTP, outbound TCP 4460) on new installs; an upgrade from 24.04 keeps timesyncd |
| utilities | GNU coreutils | Rust coreutils (uutils); cp, mv and rm are still GNU — their switch was postponed to 26.10 |
| cgroup v1 | still there | removed: old containers that need v1 won’t start |
| minimum RAM | 1 GB | 1.5 GB per the installer requirements; an installed system runs on 1 GB too |
| Docker | supported | supported from release, resolute packages in Docker’s repository |
| If… | Use |
|---|---|
| a new server, the provider offers a 26.04 image, and the stack is Docker and regular packages | 26.04: two more years of updates, no move in 2029 |
| you need a third-party repository or agent that doesn’t support 26.04 yet | 24.04 |
| the server runs old containers or software that needs cgroup v1 | 24.04 |
| a server already running 24.04 | keep it or upgrade: once 26.04.1 is out, sudo do-release-upgrade offers 26.04 on 24.04 too |
The commands in the set are the same for both versions. Where 26.04 differs, the step says so.
How to set it up: by hand, cloud-init or Ansible
| How it works | When to use | |
|---|---|---|
| by hand from the manuals | commands step by step, a check after each | the first server: you see what each command does and what breaks without it |
| cloud-init (cloud-init) | YAML in the “user data” field at creation, runs on first boot | servers get created often, or the provider supports user data |
| Ansible | a playbook run from your computer over SSH, repeatable: a second run changes nothing | more than five servers, or the setup has to be repeated and compared |
The set starts with the manual path: cloud-init and Ansible do the same steps, and figuring out why something didn’t work is easier when you know the steps.
What you get and what it costs
What you get
- Password guessing is useless. Key-only login, root closed over SSH; CrowdSec bans an address after a few attempts, and the ban covers both SSH and the sites (Docker ports bypass the input chain).
- Three ports open to the outside. Everything else is
deny incoming; services only the server itself needs listen on127.0.0.1. - Holes close themselves. Once a day
unattended-upgradesinstalls packages fromnoble-security(apt-daily-upgrade.timer, around 6:00 with up to an hour of jitter) — no login needed. - The server survives memory spikes. Swap gives the kernel time instead of an instant
Out of memory.
What it costs
- A nightly reboot. With
Automatic-Reboot, every few weeks the server reboots at 4:00 — sites are down for 1–2 minutes. Without it, a new kernel doesn’t run until you reboot yourself (unattended-upgrades). - ufw and Docker live apart.
ufw deny 8080doesn’t close a container’s published port: Docker writes its own rules around it (Docker ports bypass the input chain, The ufw firewall). - Lose the key, lose the login. Passwords are closed, and only the provider console is left. Add a second key right away.
- 26.04 is still young. The Rust versions of
sudoand coreutils behave differently in places: scripts withsudo -Esilently lose variables.
Where to start
- A new Ubuntu server: what to set up in the first hour and why
- First login to a server: updates, hostname, a sudo user, key loginUpdates, hostname, time zone, your own
sudouser with a key, key-only login. The first 20 minutes. - The ufw firewall: everything closed except SSH and webufw: everything closed except SSH and web — and why it doesn’t close Docker’s ports.
- unattended-upgrades: security updates on their own and a reboot at 4:00unattended-upgrades block by block: what installs itself, when the server reboots, and how to check.
- Swap, journal and time: three settings for a small VPSThree settings for a small VPS: a swap file, a journal size limit, clock sync.
- Docker on Ubuntu: from Docker’s repository, with a log limitFrom Docker’s repository, not snap, with container log rotation. Before the bans: CrowdSec runs in a container.
- CrowdSec against SSH brute force: a ban across the whole serverCrowdSec reads
auth.logand bans addresses that guess logins, across the whole server. - An Ubuntu server: everyday commandsServices, journal, packages, disk, network — everyday commands.
- cloud-init: a server configured the moment it’s createdThe same as a file at server creation: user, key, packages, firewall.
If it doesn’t work
After ufw enable SSH won’t let you in, new sessions hang
- Cause
- The firewall went on before the SSH port was allowed: the
deny incomingpolicy closed it too. - Fix
- Provider console (Provider console):
sudo ufw allow OpenSSHorsudo ufw disable. For next time, follow the order in The ufw firewall.
Could not get lock /var/lib/dpkg/lock-frontend
- Cause
- Right after first boot
unattended-upgradesinstalls updates on its own and holds theaptlock. - Fix
- Wait a few minutes:
ps aux | grep -E 'unattended|apt'shows who holds it. Don’t delete the lock files — that’s how the package database gets broken.
Login shows *** System restart required ***
- Cause
- The kernel or a library that can’t be swapped live was updated.
- Fix
- What needs the reboot:
cat /var/run/reboot-required.pkgs; reboot at a quiet time (Is a reboot needed).
Processes on a small VPS vanish, and dmesg shows Out of memory: Killed process
- Cause
- 1–2 GB of RAM and no swap: the kernel kills the biggest process, often the database.
- Fix
- A 1–2 GB swap file (Swap, journal and time) and memory limits for containers.
The disk filled up within a few months, though there’s little data
- Cause
- The systemd journal, Docker container logs without rotation, old kernels and images.
- Fix
sudo du -xh / --max-depth=2 | sort -h | tail; limits — Swap, journal and time and Docker on Ubuntu.
The server clock lags, TLS certificates are “not yet valid”
- Cause
- Clock sync doesn’t work: on 26.04 chrony uses NTS by default and needs outbound TCP port 4460.
- Fix
timedatectlshould saySystem clock synchronized: yes; on 26.04,chronyc sources(Swap, journal and time).