When Microsoft 365 Wobbles, Game Night Shouldn't: Dedicated Servers on Linux and Docker
After another Microsoft 365 outage, this guide explains when to stop renting per-slot game servers and run your own on Linux with Docker Compose, including costs, trade-offs, and a working Minecraft example.
When Microsoft 365 Wobbles, Game Night Shouldn't: Dedicated Servers on Linux and Docker
Last week I watched a status page do the thing status pages do: turn orange, then red, then fill up with a timeline of "we're investigating" posts. Microsoft 365 was degraded, half the internet was making jokes about it, and somewhere in a group chat a friend asked whether our Valheim world was still up. It was. It's a Compose file on a box in my basement that has never once published a status page, because there's nobody to publish it to.
That's the whole pitch, really. The Microsoft 365 outage is a 700-point HN thread because it's a reminder that everything you rent is a shared failure domain. The same week, Nvidia's leadership drama and Anthropic's IPO filing were both doing numbers, and both point the same direction: compute is being repriced, and hobby-tier hosting is downstream of that repricing. Your $12/month per-slot game server host is a company with a datacenter bill, and that bill got more expensive.
So this is the article where I tell you to stop renting slots and run the server yourself. Not because self-hosting is always better — it isn't, and I'll say where — but because for the specific workloads we're talking about, it's now dramatically easier than the last time you looked.
Why This Week Matters for a Box in Your Basement
There are two forces pushing in opposite directions and it's worth being honest about both.
The first is concentration risk. When Microsoft 365 degrades, a huge slice of the working world stops. Game server hosting isn't Microsoft, but it's the same architecture: a control panel, a shared database of customer configs, a fleet of nodes, and a billing system. When any one of those breaks, you don't get to fix it. You get to read the status page. I've been on both sides of this and I'd rather own the breakage.
The second is cost. The AI buildout has been pulling datacenter capacity, power, and GPUs into a different market. That pressure shows up as higher VPS prices, fewer cheap bare-metal deals, and hosting providers quietly trimming what "unlimited slots" means. If you pay per-slot for a Minecraft server that idles at 4% CPU, you're subsidizing someone else's capacity planning.
Meanwhile, the Linux side got better. Proton has made Linux gaming a non-issue for most of my friend group — I wrote about that trajectory here — and the tooling for running dedicated servers has consolidated around a handful of Docker images that just work. Rust 1.100 shipping is also a small signal here: reimplementations of classic game servers in Rust have been creeping out of hobby projects, and a stable toolchain that fast makes them less of a novelty.
The Managed Hosting Trap
I want to be fair to managed hosts. They handle DDoS mitigation, they have a panel your least technical friend can use, and they'll restore your world from a backup at 2am so you don't have to. That's real value.
What they also have is a per-slot pricing model that scales with player count rather than resource usage, and a data export that is usually a tarball you have to figure out yourself anyway. Once you've paid for a year of "16 slots," you've spent more than a used mini PC that would run the same world with room to spare.
The lock-in is the part that bothers me more than the price. Your world, your mod list, your plugin configs, your whitelist — they live in a database you can't query and a filesystem you access through a web UI. It's the same pattern I complained about in the egress fee piece: the fee goes away, the friction stays.
Here's how the options actually compare for something like an 8-player survival server:
| Approach | Setup effort | Realistic cost | Who fixes it when it breaks | Best for |
|---|---|---|---|---|
| Managed game host | 15 min | $10–25/mo | Their support, on their schedule | Groups with zero Linux appetite |
| VPS + LinuxGSM + systemd | 2–3 hours | $6–15/mo | You, via SSH | Tinkerers who want a familiar OS |
| VPS or home box + Docker Compose | 1–2 hours | $0–12/mo | You, via docker compose |
Most people reading this |
| Panel (Pterodactyl/AMP) hosting many servers | A weekend | $20+/mo or your own hardware | You, via the panel | Communities running 5+ games |
| Kubernetes | A week you won't get back | Whatever your cluster costs | You, at 3am, crying | People who already run k8s |
For a single server, Docker Compose wins on almost every axis. You get image-based updates, a declarative config you can commit to git, and a blast radius of exactly one container.
The Actual Build: Minecraft in About Ten Minutes
I'll use Minecraft because it's the worst case in some ways — Java, memory-hungry, needs a proper EULA flag — and if it works here it works everywhere.
Create a directory, drop in a compose.yaml, and set an RCON password in an .env file next to it:
# compose.yaml
services:
minecraft:
image: itzg/minecraft-server:java21
container_name: mc-survival
network_mode: host # see the networking section below
environment:
EULA: "TRUE"
TYPE: "PAPER"
VERSION: "1.21.5"
MEMORY: "6G"
DIFFICULTY: "hard"
VIEW_DISTANCE: "8"
SIMULATION_DISTANCE: "6"
ENABLE_RCON: "true"
RCON_PASSWORD: "${RCON_PASSWORD:?set RCON_PASSWORD in .env}"
MAX_PLAYERS: "12"
volumes:
- ./data:/data
- ./plugins:/plugins
ulimits:
nofile: 65535
restart: unless-stopped
# .env — chmod 600 this
RCON_PASSWORD=$(openssl rand -hex 24)
Then:
docker compose up -d
docker compose logs -f minecraft | grep -i 'done'
When you see Done (12.3s)! For help, type "help", connect. That's it. Updates are docker compose pull && docker compose up -d. Changing the modpack or version is a two-line edit. Rolling back a bad update is a git checkout if you keep the file in a repo.
The itzg/minecraft-server image has been maintained for years and handles the annoying parts — EULA, world upgrades, plugin downloads, RCON. I have no affiliation; I've just never had it eat a world.
Steam Games Are Slightly More Work
Valheim, CS2, Palworld, Factorio — these all come down to SteamCMD. The pattern is the same: fetch the app, mount the data, run the container on host networking.
#!/usr/bin/env bash
# /usr/local/bin/valheim-update.sh
set -euo pipefail
APPID=896660
INSTALL_DIR=/srv/valheim
/usr/games/steamcmd \
+force_install_dir "$INSTALL_DIR" \
+login anonymous \
+app_update "$APPID" validate \
+quit
Run that from a systemd timer weekly, then start the server container:
docker run -d --name valheim \
--network host \
-v /srv/valheim:/home/steam/valheim \
-v /srv/valheim/saves:/home/steam/.config/unity3d/IronGate/Valheim \
-e SERVER_NAME="Basement" \
-e WORLD_NAME="Basement" \
-e SERVER_PASS="$(cat /etc/valheim-pass)" \
-e SERVER_PUBLIC=0 \
--restart unless-stopped \
lloesche/valheim-server
Two things matter here. First, --network host — Valheim uses UDP 2456–2458, and while Docker can publish UDP fine, host networking sidesteps a pile of NAT weirdness and the userland proxy entirely. Second, SERVER_PUBLIC=0 — your world doesn't need to be on the community list for four friends to play.
The update script is where I'd point you at my home lab automation writeup, because the difference between a server you enjoy and a server you resent is whether updating it is a chore or a timer.
The Things That Actually Bite You
Nobody warns you about these, so here's the short list.
Networking. For anything with a query port, a lobby, or peer discovery, use network_mode: host. Docker's bridge networking plus the userland proxy adds a hop and a conntrack entry per connection. Set userland-proxy: false in /etc/docker/daemon.json regardless. And if you're on a busy box, check net.netfilter.nf_conntrack_max — a 64-player CS2 server can exhaust the default on some distros.
CPU, not cores. Game servers are latency-sensitive and single-threaded in the hot path. Pinning the container with cpuset: "2-3" on a 4-core VPS, and setting the CPU governor to performance on a home box, does more for tick stability than doubling RAM. Docker's --cpus is a fairness knob, not a latency knob.
Security. A game server is an internet-facing process running untrusted user input. Run it as a non-root UID inside the container (user: "1000:1000"), and never expose RCON to the world. I put admin ports behind WireGuard; RCON on 25575 feels fine until it isn't.
Backups. Snapshot the volume while the server is stopped, or use RCON's save-off/save-all before tar. A restic job to object storage costs pennies. I've lost exactly one world to a corrupted save and I think about it more than I should.
What I'd Do This Weekend
If you're currently paying per-slot, don't migrate everything at once. Pick your least-played server, export the world, and run it in Docker on a $6 VPS for a month. Keep the managed host alive as a fallback while you prove the workflow. Then cancel it and move the next one.
If you're starting fresh, skip the panel. docker compose plus a git repo is a better control plane than any web UI, and you already know how to use it.
And if you're running five games for thirty people — yeah, run Pterodactyl. That's a real use case for a panel. Kubernetes for a Minecraft server shared with four friends is a hobby, not an architecture, and I say that as someone who has absolutely done it.
Recommended tools & services
Disclosure: some links below are affiliate links. If you buy through them I may earn a commission at no extra cost to you. I only list tools I actually run.
| Tool | Why it is here | |
|---|---|---|
| Amazon - SSDs & NVMe | Boot drives and storage for your homelab | Visit |
| Amazon - Mini PCs | Low-power always-on nodes (ProDesk, EliteDesk, N100 boxes) | Visit |
| Amazon - PC Components | CPUs, RAM and cases for a build | Visit |
| Amazon - GPUs | Consumer GPUs for local inference (16GB VRAM is the sweet spot) | Visit |
| Amazon - RAM & Memory | More RAM is the cheapest upgrade for VMs and containers | Visit |
| Amazon - Game Controllers | 8BitDo and similar pads work great with RetroArch | Visit |