I’m thinking about starting a self hosting setup, and my first thought was to install k8s (k3s probably) and containerise everything.
But I see most people on here seem to recommend virtualizing everything with proxmox.
What are the benefits of using VMs/proxmox over containers/k8s?
Or really I’m more interested in the reverse, are there reasons not to just run everything with k8s as the base layer? Since it’s more relevant to my actual job, I’d lean towards ramping up on k8s unless there’s a compelling reason not to.
Unless you have multiple systems, I don’t think k8s will yield much benefit over plain docker.
So, if I plan to build a pi cluster I should get familiar with k8s?
The basics can be useful there. The whole idea with k8s is to be able to run applications across multiple hosts in a given fleet. Your cluster can be that fleet! :)
Also k8s is in high demand in the sector, so those are good skills that could be turned into $$
I get why too. I’m a full stack (including devops) software engineer, and docker/k8s is just completely opaque to me. I’m not sure why, but I really just can’t wrap my head around it. Thankfully my current company has a devops team that takes care of it, but jeez
Tbh those stuff aren’t really intuitive. But, as was my case for instance, that’s something that can be “easily” learnt as a hobbyist like us. And when you understand those concepts, at least from an abstract point, my stance is that you can become a better dev/ops/sys :) I strongly advice anyone in the field to at least play a little with Docker/containers to grasp what it is.
I can’t even get my head around Traefik, let alone Kubernetes (and k3 vs k8???)
I’m running a 3 pi cluster with k3s at the moment. The main benefit I’ve found is that all my pis run exactly the same software setup as a base so it’s easy to add new ones or replace/update one. I use a deployment management application to push my deployments too which means it’s super easy to redeploy everything if something goes funky.
That can be fun. The benefit of kubernetes is flexibility in the orchestration and (sometimes) scaling. Also the tooling in Kubernetes is more sofisticated compared to plain containers or manual services.
Kubernetes is basically just a finite-state machine that is able to manage a certain number of nodes as a pool of resources. This has added complexity compared to you managing the scheduling (I.e. I install this service on this box and this on this other box), but it also allows for much easier automation.
A multitude of things are far easier to do on Kubernetes. If you combine it with an immutable OS, then less effort too.
I think it depends on your scale. If homelab stuff docker is awesome IMO.
I, personally, haven’t done a whole lot of VM work but I do run a metric ass-ton of containers. I can spool up servers in docker compose on absolutely dogshit hardware and have it run serviceably. Also, the immutability of the container OS is really nice for moving things around and/or getting them set up quickly.
Where did you learn so much about Docker? Having a server at home, I’m more inclined to spin up a VM. I would like to learn more about Docker.
If I’m honest, I’ve stumbled nose-first through pretty much everything I know. I am never afraid to break things as long as I learn from it.
Just get started somewhere. I ran traditional VMs for most things before and I would never go back unless it was necessary for something.
Easiest way is just to start using Docker for some service you’re hosting that has a public image available and go from there. If you want a more visual approach there’s stuff like Portainer you can use too.
Also get started early on with docker compose, it makes it much easier to organize your container configs.
VMs are often imperative and can be quite easy and familiar to setup for most people, but can be harder or more time-consuming to reproduce, depending on the type of update or error to be fixed. They have their own kernel and can have window managers and graphical interfaces, and can therefore also be a bit resource heavy.
Containers are declarative and are quite easy to reproduce, but can be harder to setup, as you’ll have to work by trial-and-error from the CLI. They also run on your computers kernel and can be extremely slimmed down.
They are both powerful, depends how you want to maintain and interface with them, how resource efficient you want them to be, and how much you’re willing to learn if necessary.
That sums it up really well.
I generally tend to try to use containers for everything and only branch out to VMs if it doesn’t work or I need more separation.
This is my general recommendation as containers are easier to set up and in my opinion individual software packages are easier to maintain with things like compose. I have limited time for my self hosted instance and that took away a lot of work, especially when updating.
That sums it up really well.
I generally tend to try to use containers for everything and only branch out to VMs if it doesn’t work or I need more separation.
This is my general recommendation as containers are easier to set up and in my opinion individual software packages are easier to maintain with things like compose. I have limited time for my self hosted instance and that took away a lot of work, especially when updating.
What I did is install proxmox on the bare metal, setup a vm in which I put the containers.
Proxmox itself stays (almost) completely stock. The only changes I’ve made to it were to add the NUT client package so it could gracefully shut down if my NUT server indicates that the UPS is running out of power during an outage.
In your VMs you can do whatever. Setup OMV, or a stock Ubuntu or Debian vm and install your services on the VM or use Docker/Podman. Setup Fedora CoreOS or IoT vms and host all your services in Podman containers.
The great thing about Proxmox is you can do snapshot backups which take mere moments to complete. Then pass those off to a NAS where they can survive a irreparable loss of your Proxmox server.
You can also spin up new vms as needed to just try to fuck around with new techs or just a new way of setting up your home lab. It gives you a ton of flexibility and makes backing stuff up way easier.
Another great thing you can do is if 3 years down the line you are looking to replace your server hardware with some newer or more powerful stuff you can just add the new device as a node to the cluster. Then you can migrate all your existing VMs over to your new hardware and decommission your old one with very little to no downtime on anything.
The great thing about Proxmox is you can do snapshot backups which take mere moments to complete. Then pass those off to a NAS where they can survive a irreparable loss of your Proxmox server.
Hopefully you put a giant asterix by this point. You need the snapshot AND the original backup. Snapshots are only diffs and can’t survive without their base backup.
This is my exact setup as well. Proxmox with one beefy vm dedicated just to docker and then a few other vms for non docker workloads (eg, home assistant, pihole, jelltfin). I can probably run those in docket as well, but the to worked better as vms when I set them up
Appreciate your take on this and specifically mentioning that you have a VM for Home Assistant. That was a lightbulb moment for me as I like how easy it is to manage updates as an OS install rather than in a Docker container. If I ever get around to rebuilding my server architecture I’m definitely going to do this!
I have a similar setup, but 2 VMs on each of my 2 servers, then on server 1, I have VM A running one test K3s node and VM B running one live (Production) K3s node with the same on server 2, so I can take one server full down for maintenance, but keep my test and live sites running. It’s way overkill, but allows me to learn about how to set up and maintain resilient systems. One day, I’ll do the same for my network :-(
Container processes are just ordinary linux processes, so they don’t need extra overhead (cpu and ram reservation) to run, which means your machine can run more of them. If you have a machine with 32GB of ram, can probably run 15 VMs with 2GB of ram each where the actual app running inside the VM might only consume about 50% of the VM ram, or you can run them as container and they all would just consume 15GB of ram, leaving you extra to run more containers. I found this to be ideal for self hosting because all apps are your personal apps so interprocess isolation is not as important compared to running in public cloud.
I’ve always been unclear of why people choose to run VM’s. I would think you’d want to try Docker first, LXC second, and VM only in the last instance, if you need to emulate a different architecture? But if the stuff you need to run has been ported to your server’s architecture why add the overhead?
There’s been some nasty buggery with avahi instances on containers clashing with host ones in the past
Some programs just don’t like to run without access to parts to your system like /proc /sys and /run.
Rather than bother with crafting bespoke permissions, non-default cgroups and elevated rights for certain containers, I’ve definitely opted for just installing a VM.
It was always a time/functionality choice, and not one I make often - crafting the right solution is always better; but I have done it
It depends on your use case and what you are trying to achieve.
You do not need k8s (or k3s…) to use containers though. Plain old containers could also suffice, or Docker Swarm if you need some container orchestration functionality.
Trying to learn k8s would be a good reason to use k8s though :)
I personally really, really like (Docker) containers and I host most of my stuff with it, on a Raspberry Pi and on (free tier) Oracle Cloud VPS’s. I also plan to (re)install Proxmox on a spare old laptop and run some stuff in VMs on that (namely Home Assistant) and might try a NixOS server too.
So really, use both. Use the right tool for the job. And you can also run containers in VMs and even use Ansible to configure everything with playbooks, allowing you to re-run said playbooks when things go wrong.
Personally I always use containers unless there is a good reason to use a VM, and those reasons do exist. Sometime you want a whole, fully functional OS complete with custom kernel, in that situation a VM is a good idea, sometimes a utility only comes packaged as a VM.
But absent of a good reason, containers are just better in the majority of cases
If everything you want to run makes sense to do within k8s it is perfectly reasonable to run k8s on some bare-metal OS. Some things lend themselves to certain ways of running them better than others. E.g. Home Assistant really does not like to run anywhere but a dedicated machine/VM (at least last time I looked into it).
Regardless of k8s it may make sense to run some sort of virtualization layer just to make management easier. One panel you can use to access all of the machines in you k8s cluster from a console level can be pretty nice, and a Proxmox cluster would give you this. You can make a VM on a host that takes up basically all of the available RAM/CPU on it. Proxmox specifically has some built-in niceties with gluster (which I’ve never use, I manage gluster myself on bare metal) which could even be useful inside a k8s cluster for PVCs and the like.
If you are willing to get weird (and experimental) look into Rancher’s Harvester it’s an HCI platform (similar to Proxmox or vSphere) that uses k8s as its base layer and even manages VMs through k8s APIs… I played with it a bit and it was really neat, but opted for bare metal Ubuntu for my lab install (and actually moved from rke2 to k3s to Nomad to docker compose with some custom management/clustering over the course of a few years).
Fwiw I’ve been running home assistant in a docker container for a couple years without any issues
Yeah, I think the problem comes if you don’t want to manually configure “Add-ons”. Using this feature is only supported on their OS or using “Supervised”. “Supervised” can’t itself be in a container AFAIK, only supports Debian 12, requires the use of network manager, “The operating system is dedicated to running Home Assistant Supervised”, etc, etc.
My point is they heavily push you to use a dedicated machine for HASS.
Yeah, I haven’t missed the HASS add-ons but that’s a good point
Yea I’ve been running “core” in docker-compose and not the “supervised” or whatever that’s called.
It’s been pretty flawless tbh.
It’s running in docker-compose in a VM in proxmox.
At first, it was mostly because I wanted to avoid their implementation of DNS, which was breaking my split-horizon DNS.Honestly, once you figure out docker-compose, it’s much easier to manage than the supervised add-on thing. Although the learning curve is different.
Just the fact that your add-ons don’t need to go down when you upgrade hass makes this much easier.I could technically run non-hass related containers in that docker, but the other important stuff is already in lxc containers in proxmox.
Not everything works in containers, so having the option to spin a VM is neat.I’m also using PCI passthrough so my home theater/gaming VM has access to the GPU and I need a VM for that.
Even if they only want to use k8s or dockers for now, having the option to create a VM is really convenient.
I use proxmox for the sole benefit of just spinning up a VM of choice without having to deal with usb-sticks, etc.
From there I just run everything with Docker containers, via Portainer.
This is exactly what I do for my personal servers (except with ESXi instead of proxmox).
You will probably want both VMs and containers, there are some things that are not well supported in containers (e.g. gitlab).
I run a couple k8s clusters for work and the complexity is beyond what most people starting out would want, I would imagine.
Unless you need something that has a helm chart but not docker support (e.g. gitlab) or you are really keen on learning, it can be quite a jump…
(For gitlab I still would recommend a VM with the omnibus installer over k8s unless you are big enough to have a separate team managing your k8s clusters. It would suck to have a PV issue and lose all your data.)
Just to add my two cents: When I started out I thought I’d need a datacenter, with 10 Gig connectivity and a lot of storage. Turns out, a Raspberry Pi 4 8GB would’ve been sufficient for the things I actually use.
My recommendation would be therefore to start minimalistic and build up according to your needs from there. Start with a Raspberry PI and Docker or use a used Micro SFF and go up from there, this advice would’ve saved me a lot of money and electricity.
My backup solution is rsync and so I really like docker-compose since it usually means there is zero config for restoration of backups on a new computer besides installing docker-compose (which is usually one line on the terminal).
Basically, it’s “why not both?”
So first, kubernetes is a different ball of wax than containers, and if you want to run it on one machine you can, but it’s really for running containers across a cluster of machines. I’m guessing you just generally mean containers so I’ll go with that.
Containers are essentially just apps running on a virtual os. Virtual machines are an OS running on virtual hardware. You can abstract both layers and have virtual hardware running an os that runs a virtual os for your containers, and nothing will really mind - in fact that’s kind of the way to do it if you have one big machine you need to run a bunch of services on. You might cut up a server into a Linux VM, a Windows VM, and a BSD VM, and run containers on each one. Or you might run 3 Linux VMs and have the containers for 3 different services split between them.
It really depends on what you’re hosting and trying to do for how exactly to go about it. Take for instance a pretty common self hosted stack:
Plex Radarr Prowlarr Deluge TrueNAS
Now you could install TrueNAS scale and run all of those as containers on it, and it would work ok, but TrueNAS scale isn’t really meant for managing a ton of containers right now. You could make a vm on it for each service and have them all talk to each other but then you’re probably wasting resources by duplicating the OS 5 times. Also, what if you want to run TrueNAS core instead of scale? Can you get everything else working in jails – maybe? – but it’ll probably be a pain.
Instead, you might install proxmox and pass through the drive controller, and set up one VM for TrueNAS core. Then you might make another VM for the arrs containers, and a third for Plex itself.
It gets you the best of both worlds. TrueNAS can run on BSD instead of Linux, your arrs are easy to deploy and update in containers that keep everything separated, and Plex is sequestered in a hardened os with read only access to everything else since it gets a port forwarded and is more of a security risk. Again that’s just one option though.
VMs get you a ton of really handy things like snapshots and for simple VMs, very easy portability between relatively similar hardware. I’ll probably get ruined for saying this but they’re also a security tool that you should probably keep in your belt. If someone manages to break out of a container and your files are just sitting there for the taking that’s not great. If someone manages to break into your VM and “the good stuff” is on another VM that’s another layer of security they have to break through.
Containers on the other hand use way fewer resources, especially ram - and are much easier to wrangle than many OSes for updates and config.
There’s really a lot of self hosted stuff that assumes you’re running docker and treats regular install as a kind of weird edge case, so you’ll probably run docker even if you don’t want to.
Kubernetes on the other hand I would argue isn’t really meant for self hosting where you probably have a one or two servers that you own. Its meant to deploy containers across various cloud servers in a way that’s more automated to manage. If you need storage in a kubernetes cluster you’ll probably use something like s3 buckets, not a hard drive.
If you want to learn it you can totally deploy it on a computer running a few VMs as nodes or with a few laptops / SBCs as a cluster, but if you just want the services to run on your server in the closet it’s a bit like using a sledgehammer to nail a chair back together. That’s why you don’t tend to see it talked about as much - it’s a bit of a different rabbit hole.
Containers, unless you have a specific need for a VM.
With a VM you have to reserve resources exclusively. If you give a VM 2gb of ram, then that’s 2gb of ram that you can’t use for other things, even if the guest OS is using less.
With Containers, you only need as many resources as the process inside the container requires at the time.
