Hello! 😀
I want to share my thoughts on docker and maybe discuss about it!
Since some months I started my homelab and as any good “homelabing guy” I absolutely loved using docker. Simple to deploy and everything. Sadly these days my mind is changing… I recently switch to lxc containers to make easier backup and the xperience is pretty great, the only downside is that not every software is available natively outside of docker 🙃
But I switch to have more control too as docker can be difficult to set up some stuff that the devs don’t really planned to.
So here’s my thoughts and slowly I’m going to leave docker for more old-school way of hosting services. Don’t get me wrong docker is awesome in some use cases, the main are that is really portable and simple to deploy no hundreds dependencies, etc. And by this I think I really found how docker could be useful, not for every single homelabing setup, and it’s not my case.
Maybe I’m doing something wrong but I let you talk about it in the comments, thx.
I’m actually doing the opposite :)
I’ve been using vms, lxc containers and docker for years. In the last 3 years or so, I’ve slowly moved to just docker containers. I still have a few vms, of course, but they only run docker :)
Containers are a breeze to update, there is no dependency hell, no separate vms for each app…
More recently, I’ve been trying out kubernetes. Mostly to learn and experiment, since I use it at work.
And I’ve done the exact opposite moves everything off of lxc to docker containers. So much easier and nicer less machines to maintain.
Less “machines” but you need to maintain docker containers at the end
For me the power of docker is its inherent immutability. I want to be able to move a service around without having to manual tinker, install packages and change permissions etc. It’s repeatable and reliable. However, to get to the point of understanding enough about it to do this reliably can be a huge investment of time. As a daily user of docker (and k8s) I would use it everyday over a VM. I’ve lost count of the number of VMs I’ve setup following installation guidelines, and missed a single step - so machines that should be identical aren’t. I do however understand the frustration with it when you first start, but IMO stick with it as the benefits are huge.
Yeah docker is great for this and it’s really a pleasure to deploy apps so quickly but the problems comes later, if you want to really customize the service to you, you can’t instead of doing your own image…
In most cases you can get away with over mounting configuration files within the container. In extreme cases you can build your own image - but the steps for that are just the changes you would have applied manually on a VM. At least that image is repeatable and you can bring it up somewhere else without having to manually apply all those changes in a panic.
Are you using docker-compose and local bind mounts? I’d not, you’re making backing up uch harder than it needs to be. Its certainly easier than backing up LXCs and a whole lot easier to restore.
i’m using all of this yeah
Docker is a convoluted mess of overlays and truly weird network settings. I found that I have no interest in application containers and would much prefer to set up multiple services in a system container (or VM) as if it was a bare-metal server. I deploy a small Proxmox cluster with Proxmox Backup Server in a CT on each node and often use scripts from https://community-scripts.github.io/ProxmoxVE/. Everything is automatically backed up (and remote sync’d twice) with a deduplication factor of 10. A Dockerless Homelab FTW!
Yeah I share your point of view and I think I’m going this way. These scripts are awesome but I prefer writing mine as I get more control over them
Docker compose plus using external volume mounts or using the docker volume + tar backup method is superior
Can be but I’m not enough free, and this way I run lxc containers directly onto proxmox
You’re basically adding a ton of overhead to your services for no reason though
Realistically you should be doing docker inside LXC for a best of both worlds approach
I accept the way of doing, docker or lxc but docker in a lxc is not suitable for me, I already tried it and I’ve got terrible performance
Just run docker in an LXC. That’s what I do when I have to.
Not working good on my side, performance issues
I’ve never really like the convoluted docker tooling. And I’ve been hit a few times with a docker image uodates just breaking everything (looking at you nginx reverse proxy manager…). Now I’ve converted everything to nixos services/containers. And i couldn’t be happier with the ease of configuration and control. Backup is just.a matter of pushing my flake to github and I’m done.
Already said but I need to try NixOS one day, this thing seems to worth it
Are you using docker compose scripts? Backup should be easy, you have your compose scripts to configure the containers, then the scripts can easily be commited somewhere or backed up.
Data should be volume mounted into the container, and then the host disk can be backed up.
The only app that I’ve had to fight docker on is Seafile, and even that works quite well now.
using docker compose yeah. I find hard to tweak the network and the apps settings it’s like putting obstacles on my road
Its networking is a bit hard to tweak, but I also dont find I need to most of the time. And when I do, its usually just setting the network to host and calling it done.
Docker as a technology is a misguided mess but it is an effective tool.
Podman is a much better design that solves the same problem.
Containers can be used well or very poorly.
Docker makes it easy to ship something without knowing anything about System Engineering which some see as an advantage, but I don’t.
At my shop, we use almost no public container images because they tend to be a security nightmare.
We build our own images in-house with strict rules about what can go inside. Otherwise it would be absolute chaos.
Cool I don’t want to know about system engineering and if they is your requirement to use software then nobody would be using it.
I use podman using home-manager configs, I could run the services natively but currently I have a user for each service that runs the podman containers. This way each service is securely isolated from each other and the rest of the system. Maybe if/when NixOS supports good selinux rules I’ll switch back to running it native.
This sounds great! I’d love to see your config. I’m not using home manager, but have 1 non root user for all podman containers. 1 user per service seems like a great setup.
Yeah it works great and is very secure but every time I create a new service it’s a lot of copy paste boilerplate, maybe I’ll put most of that into a nix function at some point but until then here’s an example n8n config, as loaded from the main nixos file.
I wrote this last night for testing purposes and just added comments, the config works but n8n uses sqlite and probably needs some other stuff that I hadn’t had a chance to use yet so keep that in mind.
Podman support in home-manager is also really new and doesn’t support pods (multiple containers, one loopback) and some other stuff yet, most of it can be compensated with the extraarguments but before this existed I used pure file definitions to write quadlet/systemd configs which was even more boilerplate but also mostly copypasta.Gaze into the boilerplate
{ config, pkgs, lib, ... }: { users.users.n8n = { # calculate sub{u,g}id using uid subUidRanges = [{ startUid = 100000+65536*( config.users.users.n8n.uid - 999); count = 65536; }]; subGidRanges = [{ startGid = 100000+65536*( config.users.users.n8n.uid - 999); count = 65536; }]; isNormalUser = true; linger = true; # start user services on system start, fist time start after `nixos-switch` still has to be done manually for some reason though openssh.authorizedKeys.keys = config.users.users.root.openssh.authorizedKeys.keys; # allows the ssh keys that can login as root to login as this user too }; home-manager.users.n8n = { pkgs, ... }: let dir = config.users.users.n8n.home; data-dir = "${dir}/${config.users.users.n8n.name}-data"; # defines the path "/home/n8n/n8n-data" using evaluated home paths, could probably remove a lot of redundant n8n definitions.... in { home.stateVersion = "24.11"; systemd.user.tmpfiles.rules = let folders = [ "${data-dir}" #"${data-dir}/data-volume-name-one" ]; formated_folders = map (folder: "d ${folder} - - - -") folders; # a function that takes a path string and formats it for systemd tmpfiles such that they get created as folders in formated_folders; services.podman = { enable = true; containers = { n8n-app = { # define a container, service name is "podman-n8n-app.service" in case you need to make multiple containers depend and run after each other image = "docker.n8n.io/n8nio/n8n"; ports = [ "${config.local.users.users.n8n.listenIp}:${toString config.local.users.users.n8n.listenPort}:5678" # I'm using a self defined option to keep track of all ports and uids in a seperate file, these values just map to "127.0.0.1:30023:5678", a caddy does a reverse proxy there with the same option as the port. ]; volumes = [ "${data-dir}:/home/node/.n8n" # the folder we created above ]; userNS = "keep-id:uid=1000,gid=1000"; # n8n stores files as non-root inside the container so they end up as some high uid outside and the user which runs these containers can't read it because of that. This maps the user 1000 inside the container to the uid of the user that's running podman. Takes a lot of time to generate the podman image for a first run though so make sure systemd doesn't time out environment = { # MYHORSE = "amazing"; }; # there's also an environmentfile option for secret management, which works with sops if you set the owner of the secret/secret template extraPodmanArgs = [ "--pull=newer" # always pull newer images when starting, I could make this declaritive but I haven't found a good way to automagically update the container hashes in my nix config at the push of a button. ]; # few more options exist that I didn't need here }; }; }; }; }
I used docker for my homeserver for several years, but managing everything with a single docker compose file that I edit over SSH became too tiring, so I moved to kubernetes using k3s. Painless setup, and far easier to control and monitor remotely. The learning curve is there, but I already use kubernetes at work. It’s way easier to setup routing and storage with k3s than juggling volumes was with docker, for starters.
What are really the differences between docker and kubernetes?
Both are ways to manage containers, and both can use the same container runtime provider, IIRC. They are different in how they manage the containers, with docker/docker-compose being suited for development or one-off services, and kubernetes being more suitable for running and managing a bunch of containers in production, across machines, etc. Think of kubernetes as the pokemon evolution of docker.
…a single compose file?!
Several services are interlinked, and I want to share configs across services. Docker doesn’t provide a clean interface for separating and bundling network interfaces, storage, and containers like k8s.
Did you try portainer?
I did come across it before, but it feels like just another layer of abstraction over k8s, and with a smaller ecosystem. Also, I prefer terminal to web UI.
Fair. It does make bundling networks easy though.
Isn’t it more effort to setup kubernetes? At work I also use k8s with Helm, Traefik, Ingress but we have an infra team that handles the details and I’m kind of afraid of having to handle the networking etc. myself. Docker-compose feels easier to me somehow.
Setting up k8s with k3s is barely two commands. Works out of the box without any further config. Heck, even a multi-node cluster is pretty straightforward to setup. That’s what we’re using at work.
I love docker, and backups are a breeze if you’re using ZFS or BTRFS with volume sending. That is the bummer about docker, it relies on you to back it up instead of having its native backup system.
What are you hosting on docker? Are you configuring your apps after? Did you used the prebuild images or build yourself?
I use the *arr suite, a project zomboid server, a foundry vtt server, invoice ninja, immich, next cloud, qbittorrent, and caddy.
I pretty much only use prebuilt images, I run them like appliances. Anything custom I’d run in a vm with snapshots as my docker skills do not run that deep.
This why I don’t get anything from using docker I want to tweak my configuration and docker is adding an extra level of complexity
Tweak for what? Compile with the right build flags been there done that not worth the time.
If I want really to dive in the config files and how this thing works, no normal install I can really easily, on docker it’s something else
What application are you trying to tweak?
I should also say I use portainer for some graphical hand holding. And I run watchtower for updates (although portainer can monitor GitHub’s and run updates based on monitored merged).
For simplicity I create all my volumes in the portainer gui, then specify the mount points in the docker compose (portainer calls this a stack for some reason).
The volumes are looped into the base OS (Truenas scale) zfs snapshots. Any restoration is dead simple. It keeps 1x yearly, 3x monthly, 4x weekly, and 1x daily snapshot.
All media etc… is mounted via NFS shares (for applications like immich or plex).
Restoration to a new machine should be as simple as pasting the compose, restoring and restoring the Portainer volumes.
I don’t really like portainer, first their business model is not that good and second they are doing strange things with the compose files
I’m learning to hate it right now too. For some reason, its refusing to upload a local image from my laptop, and the alarm that comes up tells me exactly nothing useful.
Docker is good when combined with gVisor runtime for better isolation.
What is gVisor?
gVisor is an application kernel, written in memory safe Golang, that emulates most system calls and massively reduces the attack surface of the kernel. This is important since the host and guest share the same kernel, and Docker runs rootful. Root inside a Docker container is the same as root on the host, as long as a sandbox escape is used. This could arise if a container image requires unsafe permissions like Docker socket access. gVisor protects against privilege escalation by only using root at the start and never handing root over to the guest.
Sydbox OCI runtime is also cool and faster than gVisor (both are quick)
I like reminding people that with every new technology, the old one is still around. The new gets most of the attention, but the old is still kicking. (We still have wire wrapped programs kicking around.)
You are all good. Spend your limited attention on other things.
Yeah, when I got started I initially put everything in Docker because that’s what I was recommended to do, but after a couple years I moved everything out again because of the increased complexity, especially in terms of the networking, and that you now have to deal with the way Docker does things, and I’m not getting anything out of it that would make up for that.
When I moved it out back then I was running Gentoo on my servers, by now it’s NixOS because of the declarative service configuration, which shines especially in a server environment. If you want easy service setup, like people usually say they like about Docker, I think it’s definitely worth a try. It can be as simple as “services.foo.enable = true”.
(To be fair NixOS has complexity too, but most of it is in learning how the configuration language which builds your operating system works, and not in the actual system itself, which is mostly standard except for the store. A NixOS service module generates a normal systemd service + potentially other files in the file system.)
nixos definitely gives a try
I ditched nix and install software only through portage. If needed, i make my own ebuilds.
This has two advantages:
- it removes all the messy software: i am not going to install something if I can’t make the ebuild becayse the development was a mess , like everything TS/node
- i can install, rollback, reinstall, upgrad and provision (configuration) everything using portage
- i am getting to know gentoo and portage in great details, making the use of my desktop and laptop much much easier
