Everything I self-host at home
tl;dr
I self-host Nextcloud, Immich, Jellyfin, Home Assistant, Invidious and a few supporting tools on three machines behind a single Caddy entrypoint. Nightly updates, Gatus + ntfy monitoring and a manual unlock chain for encrypted disks keep it running, for ~50 EUR/month of avoided subscriptions against ~15 EUR/month of costs.
I didn’t start self-hosting out of frustration with the cloud. I started because I wanted to understand what I rely on, and to keep my photos, files and media under my own control.
Over time it became a small home setup that stays stable, cuts some recurring costs, and supports the projects I work on.
The hardware
Three machines, one role each:
illustrious | formidable | portland | |
|---|---|---|---|
| Role | public entrypoint | workstation + compute | storage |
| OS | Arch Linux (kernel 6.17.11-hardened) | CachyOS (kernel 6.12.63-cachyos-lts) | Arch Linux (kernel 6.12.59-lts) |
| CPU | Intel Core i5-1135G7 (8) @ 4.20 GHz | AMD Ryzen 9 7950X (32) @ 5.88 GHz | AMD Ryzen 7 2700X (16) @ 3.70 GHz |
| GPU | NVIDIA RTX 4090 | ||
| RAM | 31.00 GiB | 61.91 GiB | 31.26 GiB |
| Disks | 914.83 GiB (ext4) 915.82 GiB (ext4) 3.64 TiB (xfs) | 929.50 GiB (btrfs) 1.82 TiB (xfs) 7.28 TiB (xfs) 119.24 GiB (btrfs) | 697.49 GiB (btrfs) 3.64 TiB (fuseblk) 3.51 TiB (zfs) |
The operating systems follow the roles. The always-on machines run Arch Linux for its straightforward tooling and the control it gives me, along with the maintenance it asks for. formidable is my daily workstation and runs CachyOS, because its performance-focused defaults keep it responsive without constant tuning.
One public entrypoint
Everything is deployed with Docker Compose. It keeps the setup consistent, repeatable, and easy to rebuild.
Only one machine faces the internet: illustrious. My domain points at it, and it’s the only place my router forwards ports 80/443 to. Caddy is the reverse proxy: it handles HTTPS and routes requests to the right containers, including services that actually live on portland.
That makes illustrious a single point of failure, and I accept it. The goal is reliability and simplicity, not a full high-availability replica.
For remote access, I don’t expose SSH: it’s VPN-only, keys only, no password prompt. A few services are IP-locked too, and authentication is handled per service. Not fancy, but consistent with the rest: fewer moving parts, fewer surprises.
What I actually host
The easiest way to explain the stack is by the small frictions it removed from my days.
Media
For my own library, Jellyfin is the media server. The collection is DVD rips, Blu-ray rips, CDs and music I bought. It’s a few terabytes, and the point is reliable access to my own media, not a broad streaming catalogue.
I don’t transcode. I mostly direct-play on mobile, Apple TV and the web. That one decision makes the whole thing simpler and sturdier: if a file plays, it plays. If it doesn’t, I fix the file, not the server.
For YouTube, I use Invidious, an open-source alternative front-end. It’s the most fragile part of the stack, not because the server is unstable, but because YouTube changes often and upstream breakage is unavoidable. When it breaks, I update and move on.
Photos and files for the family
The backbone for family storage and sync is Nextcloud. In practice we mostly use it as a shared drive. It’s multi-user, reliable and predictable, which is exactly what I need.
For photos and videos, I run Immich. The library is big (100k+ photos and videos), several phones back up to it, and I use the ML features. The always-on servers don’t do the heavy lifting: ML indexing is delegated to formidable, because that’s literally what a 4090 is for when it’s not being dramatic about drivers. (kinda outdated joke nowadays)
Home automation
Home Assistant is there for practical reasons: Tapo smart plugs to measure electricity use, and turning devices on or off remotely. It shows where power goes and cuts unnecessary idle draw.
Monitoring and notifications
The most valuable tools are the ones that catch slow, silent failure.
Gatus checks service health and the mounted disks on illustrious and portland, including ZFS status. It tells me clearly when something drifts from expected.
For notifications I run ntfy. It covers long-running jobs and scripts where I only need “done” or “failed”. No dashboard, just a notification.
For recipes, I self-host Tandoor Recipes: no ads, no bloat, fast access to the recipes. On iOS I use kitshn (also open source), which plugs into the same library.
Experiments
On illustrious I also run bitmagnet behind Gluetun with Proton VPN. It barely affects the rest of the stack, and it stays isolated behind the VPN because of the ecosystem it touches.
Now and then, illustrious also runs a Minecraft server, which temporarily bumps RAM usage.
Updates and maintenance
I want the always-on machines secure without maintenance becoming a hobby.
- Nightly updates. On
illustriousandportland, host updates run from asystemdtimer. I track security-relevant updates aggressively (microcode, firmware, kernel, bootloader), but keep operations conservative: LTS kernel, and reboots as a planned event, not something that happens whenever. - Live patching.
kernel-livepatchlets the system absorb many kernel security fixes without an immediate reboot. I still reboot about once a week to keep things clean and predictable. - No snapshots before upgrades. I don’t run pre-flight checks or automatic snapshots. I rely on fast detection and simple intervention instead:
pacmankeeps two package versions around for downgrades, and the previous kernel is always available at boot.
The safety net is monitoring. Gatus checks ZFS pool status, mountpoints, disk space and core service health, and alerts go to ntfy after a short delay (about 3 minutes). If something breaks, I find out quickly and fix it by hand. There’s no automated rollback pipeline. The loop is update → verify → intervene, with an informal goal of being back to normal within about an hour.
I don’t have a UPS, so a power cut can still mean hands-on work (especially since encryption is deliberate friction here), but monitoring makes the failure obvious and quick to deal with.
Encryption, and the unlock chain
portland’s storage is encrypted ZFS. I use OpenZFS because family data needs snapshots and strong safeguards. The disks on illustrious are encrypted too, and formidable is fully encrypted.
So after a power cut, there’s a manual unlock step, through a chain of trust at home:
I can’t fully automate that without breaking the security model, so I settled on a compromise: an iOS Shortcut that makes it fast, but still deliberate. It’s an inconvenience, but it clearly improves my risk profile, so it stays.
Costs and savings
Before this setup was stable, we paid for a family cloud bundle, premium music and a higher streaming tier. Self-hosting doesn’t replace all entertainment spending, but it cut the number of subscriptions that felt required.
My rough numbers:
- about 50 EUR/month of third-party services avoided,
- about 9 EUR/month of electricity for the two always-on machines,
- a bit over 200 EUR upfront for
portland’s storage, bought specifically for this, which over 36 months is about 5.6 EUR/month.
The hidden cost is the temptation to grow storage beyond the original plan.
Per service, again rough and specific to my situation. Storage and photos are where self-hosting makes the biggest difference:
| Self-hosted | Typical paid alternative | Cost avoided (rough, my case) | Self-host cost (rough) |
|---|---|---|---|
| Nextcloud | OneDrive Family, Google Drive, Dropbox | ~10 to 15 EUR/month | Shared infra (storage-heavy) |
| Immich | Google Photos, iCloud Photos | ~5 to 15 EUR/month | Shared infra (storage + ML indexing on formidable) |
| Jellyfin | Netflix / Plex-style streaming | ~15 to 25 EUR/month (depends what you cancel) | Shared infra (mainly disks) |
| Invidious | YouTube Premium (or “pay with ads”) | 0 EUR/month if you did not pay Premium | Tiny compute cost |
| Home Assistant | Vendor hubs + cloud subscriptions | Usually small, but adds up | Tiny compute cost |
| Tandoor Recipes | Recipe app subscriptions | 0 to ~5 EUR/month | Tiny compute cost |
| Gatus | UptimeRobot-style monitoring | 0 to ~10 EUR/month | Tiny compute cost |
| ntfy | Push/alert services | 0 to ~5 EUR/month | Tiny compute cost |
What it gave me besides money
The main benefit isn’t the savings, it’s control and flexibility.
Having my own media and photo datasets let me build recommendation workflows for music, videos and images on top of my Jellyfin library, based on a real collection that reflects what I actually use. Self-hosting keeps that data local and available for the work.
It also made me learn infrastructure the practical way. ZFS took real time and study before I trusted it, but once it clicked, it changed how I think about storage and durability.
And it connected me with a community that treats storage and preservation with the same priorities.
Pros and cons
The biggest advantage is depending less on other people’s decisions: fewer surprise price changes, fewer retired features, fewer forced migrations, and better privacy because the data stays local. The second is speed and clarity: local access is fast, and workflows stay consistent because I control the defaults.
The main downside is that illustrious is a single point of failure, on purpose, since it’s the only public entrypoint. Invidious is fragile because YouTube changes often. And encryption adds friction during outages, which I accept for the security it buys.
If you want to try
Don’t start with ten services. Start with one that removes a real pain: files (Nextcloud), photos (Immich) or media (Jellyfin). Then add a reliability tool like Gatus so you can see the system as it grows.
You don’t need a rack. You don’t need Kubernetes. You mostly need a small box that stays on, and the patience to fix one thing at a time.
Self-hosting, for me, isn’t a rebellion against the cloud. It’s a practical way to keep core services predictable, keep sensitive data local, and support the projects I care about without depending on changing third-party policies. It’s not for everyone, but it matches how I work and how much control I want over my own infrastructure.