Being a noob and all I was wondering whats the real benefit of having a monolithic lets say proxmox instance with router, DNS, VPN but also home asssistant and NAS functionalitiy all in one server? I always thought dedicated devices are simpler to maintain or replace and some services are also more critical than others I guess?
Pretty much the tradeoff that you said. Harder to maintain an all in one box since things conflict with each other. That said, it’s also harder to maintain 10 devices instead of 2. Usually, you want to segregate your services based on maintenance schedule. Something that you reboot once a year like your router probably shouldn’t be on the same device as something that you might reboot every day, like home assistant, if you value your sanity.
Also, virtualization is pretty much dead-end now and will just make your life harder.
In terms of the easiest software available for self hosting, I would use a dedicated router and a dedicated nas, as those are fairly standalone and can be purchased as appliances. Then I would use a single machine with Debian or NixOS, and use it as a Kubernetes or Docker host. (Kubernetes is super easy with k3s and easier to maintain than Docker, but there’s a higher barrier to entry as you’d have to write your services with Pod files instead of docker-compose files)
I wouldn’t recommend something that tries to do everything, like Unraid, TrueNAS, or Proxmox, as they honestly obfuscate things and make things harder to maintain. Though they can be nice for DIY NASes.
If you’re interested in high availability and clustering for a DIY NAS, you could even look into ceph/rook, which is what I’m using for my NAS, but it’s like 20x the effort of just having a standard NFS appliance.
I don’t think I’ve ever heard anyone say this… Kubernetes is a massive pain in the ass to learn, maintain and troubleshoot. If you find it easy that’s great, but it’s not for everyone.
I mean that with k3s you can get a kubernetes cluster running with 0 effort on a single machine. It is easier to maintain, because it handles restarting containers, updating containers, managing ports, provisioning storage, creating databases, etc for you. I’ve found the logs and events system to be super useful for troubleshooting compared to Dockerd, but maybe it can be tricky if it does something you don’t expect it to.
Obviously you need to learn how to use that automation to take advantage of it, and stuff like networking and persistent volumes can be confusing if you don’t have a good guide on it. The fact that there are different drivers for networking, storage, database management, etc can also take a bit of time. That said, networking and storage can be confusing on Docker too if you don’t have a good guide, and Docker-compose also has a learning curve, so I honestly don’t think Kubernetes is that much more effort. The main thing is that most guides are written for Docker, but the Kubernetes documentation is really good too.
If you just want to just run containers for jellyfin and home-assistant, Docker compose will be good enough. But if you want databases, reverse proxy, certificates, dns, self-healing, etc, for running bigger stuff like nextcloud and lemmy, then I would spend the extra 50% effort and do it on Kubernetes, it’ll save you time and headaches in the long run.
Asking an LLM like Lllama or ChatGPT might be a good way to learn the basics with Kubernetes, but things move fast once you start getting into the newest operators like CNPG and Gateway API.
I do all that with docker… I fail to see what Kubernetes adds to that on a single machine.
Kubernetes does it a lot better. No more messing with caddy config files, or docker sockets, you get the real deal, production stuff.
Containers automatically take themselves off the built-in loadbalancer and/or restart when they fail a health check.
A new high-availability postgres cluster with automatic backups is just a Cluster, a firewall rule is just a NetworkPolicy, a new subdomain is just an HTTPRoute, a new proxy container is just a Gateway, a new auto-renewed Let’s Encrypt certificate is just a Certificate, and DNS is set up automatically with the domain name from the HTTPRoute without me touching anything. Everything is high-availability and self-healing, I’ve never had anything go down or crash.
The other thing is ArgoCD, which automatically syncs your cluster with git. If I edit any of my config files in git, it is instantly updated on the cluster itself.
Here is my configuration for my 200+ containers, even my Lemmy instance is running here: https://codeberg.org/jlh/h5b/src/branch/main/argo/custom_applications
Docker and the Docker ecosystem copies a lot of features from Kubernetes, because they’re essentially the same thing, but Kubernetes does it in a production-ready, maintainable way. Kubernetes is an automation tool that lets 1 engineer do the work of 10.
Right, right, you just have to reinvent a dozen wheels, use only software that Kubernetes knows how to work with, and learn a bunch of new names for everything.
Once you learn it it isn’t super crazy but takes a lot of effort obviously. I think most people who do use k3s and k8s at home are people who use it for work so already knows how and where things should work and be. That said I work with kubernetes every day for work managing a handful of giant production clusters and at home I use unraid to keep it simple.
deleted by creator
Yep, I think there’s sound arguments for separating out your storage (NAS) and network (router/DNS/PiHole) infrastructure. After that, whatever suits your purpose. I virtualise all my serious services on one machine under Proxmox (mostly for ease of snapshots) then have another machine for things I’m fiddling with, usually again under Proxmox so they are easy to move to production when I’m happy with them.
Makes sense. I would probably recommend more infrastructure-as-code workflows over snapshots, like ArgoCD or docker-compose, as git commits are simpler than VM snapshots. But both ways work.
I’m running an Unraid server. You can pop in and manage everything with the CLI like you would on traditional server OSes and it’ll show your containers, images, orphans etc. in the GUI and throws alerts out of the box for utilization thresholds and power events. It’s quite nice at a glance and gets the fuck out of the way the moment it’s time to be a sysadmin.
Unraid brings some good things to the table, I wouldn’t discount it completely.
I’m well aware.
I was #8 on this list: https://web.archive.org/web/20240221094039/unraid.net/about
The way that Unraid manages Docker containers is really dumb, and it gets in your way SO MUCH. Orphans are not a normal Docker idea, it is something invented by Unraid. It actively makes managing containers harder, as there is no documented way to restore orphans if I recall correctly. Creating new containers is confusing and uses non-standard terminology, when docker-compose files have been standard for half a decade now. Unraid is a really bad container orchestrator with bad abstractions and no ability to do Infrastructure as Code. The only good thing is the GUI for monitoring containers.
The monitoring GUI is nice, and I guess if you’re doing everything with the CLI and just using the GUI for monitoring it makes sense. But CLI is not a supported workflow with Unraid, and what are you paying $3/month for if you’re just going to use the CLI? I personally wouldn’t recommend the overhead, setup, and upgrade headaches over just doing the CLI with Debian. There are just as nice free dashboards available for Kubernetes.
For what it’s worth, this is my homelab: https://codeberg.org/jlh/h5b
I run nearly 300 containers in a 4 node cluster, with a separate router and iot server. Every single piece is implemented in code, because that’s easier to maintain and document. I used Proxmox for VMs/LXC for a while, and I used FreeNAS for ZFS+NFS for a while, but now I use purely NixOS and Kubernetes. I have never seen Unraid as a valuable thing that I would like to add to my homelab in the past 8 years.
Orphans are just dangling objects, are they not?
I’m only using the Unraid Docker GUI to send me utilization alerts and notify me when my images are egregiously out of date. I saw someone trying to author a compose file using the GUI once and I closed the window before the headache started.
I’m not paying $3/mo. Where’d you get that idea? I think I paid $20 for a license like 6 years ago.
I picked Unraid because I had a bunch of disparate HDDs sitting around and their filesystem intrigued me. (0 data loss after 3 drive failures so far.)
Fair enough. I think it’s bad to invent new words for “stopped container”, though. And there should be a way to re-start them.
Yeah, the container creation GUI is a mess. The $3/month thing is a new thing they started for new customers this year. https://unraid.net/pricing
Not a big deal for grandfathered users, but I think its important to consider as a new customer, as you won’t even get security updates without paying the subscription fee. Even for vulnerabilities like the CVE-2024-21626 Leaky Vessels vulnerability.
The raid is nice, but it can be kinda clunky adding/removing drives sometimes and I’ve managed to accidentally destroy an array when I was playing with it. I think you can get identical features using LVM, but obviously it’s nice how Unraid does it all for you in a GUI.
You’re not wrong!
The cheapest option Is the monthly one for no security updates, there are still regular pro and higher plans which are one and done, no grandfathering
Interesting, self hosting crossplane. What do you use it for?
Yeah I’ve been wanting to start using it. I have a colleague who uses it for platform engineering and it’s supposed to be amazing. I was going to use it for creating offsite backup buckets on OVH but I ended up setting up a Hetzner storage box manually instead because that was cheaper. Since everything I have is self hosted, really the only external infrastructure I have is Cloudflare, but all the records there are handled by external-dns, so I haven’t really seen a need to GitOps it.
One thing I do want to look at was the custom CRD feature they were talking about at Kubecon this year, it sounded like they might have finally fixed the platform engineering abstraction problem that people have been trying to use helm (badly) for. Many companies have actually been resorting to operators for this problem, which is super overkill. I did try to use cdk8s for abstraction last year, and I was even planning to create and support a production-ready Lemmy deployment option using cdk8s, but cdk8s honestly was quite clunky on the developer side and committed the sin of reimplementing an API without even properly documenting the new API.
I’m probably just going to create a Lemmy Helm chart at some point using Cloud-Native Postgres operator and Gateway API when I have time. But Helm has glaring issues, both as a developer and as a user.