I like Incus a lot, but it’s not as easy to create complex virtual networksnas it is with proxmox, which is frustrating in educational/learning environments.

This is untrue, proxmox is not a wrapper around libvirt. It has it’s own API and it’s own methods of running VM’s.

Yes, this is where docker’s limitations begin to show, and people begin looking at tools like Kubernetes, for things like advanced, granular control over the flow of network traffic.

Because such a thing is basically impossible in Docker AFAIK. You’re getting these responses (and in general, responses like those you are seeing) appear when the thing a user is attempting to do is anywhere from significantly non trivial to basically impossible.

An easy way around this, if you still want to use Docker, is addressing the below bit, directly:

no isolation anymore, i.e qbit could access (or at least ping) to linkwarden’s database since they are all in the same VPN network.

As long as you have changed the default passwords for the databases and services, and kept the services up to date, it should not be a concern that the services have network level access to eachother, as without the ability to authenticate or exploit eachother, there is nothing that they can do, and there are no concerns.

If you insist on trying to get some level of network isolation between services, while continuing to use Docker, your only real option is iptables* rules. This is where things would get very painful, because iptables rules have no persistence by default, and they are kind of a mess to deal with. Also, docker implements their own iptables setup, instead of using standard ones, which result in weird setups like Docker containers bypassing the firewall when they expose ports.

You will need a fairly good understanding of iptables in order to do this. In addition to this, if you decide this in advance, I will warn you that you cannot create iptables rules based on ip addresses, as the ip addresses of docker containers are ephemeral and change, you must create rules based on the hostnames of containers, which adds further complexity as opposed to just blocking by ip. EDIT: OR, you could give your containers static ip addresses.

A good place to start is here. You will probably have to spend a lot of time learning all of the terminology and concepts listed here, and more. Perhaps you have better things to do with your time?

*Um, 🤓 ackshually it’s nftables, but the iptables-nft command offers a transparent compatibility layer enabling easier migrations from the older and no longer used iptables

EDIT: And of course nobody has done this before and chatgpt isn’t helpful. These problems are the kinds of problems where chatgpt/llm’s begin to fall apart and are completely unhelpful. Just “no you’re wrong” over and over again as you have to force your way through using actual expertise.

You can block traffic to a Docker container by its hostname using iptables, but there’s an important nuance: iptables works with IP addresses, not hostnames. So you’ll first need to resolve the container’s hostname to its IP address and then apply the rule.

You’re right—container IPs change, so matching a single IP is brittle. Here are robust, hostname-friendly ways to block a container that keep working across restarts.

Exactly — good catch. The rule: sudo iptables -I DOCKER-USER 1 -m set --match-set blocked_containers dst -j DROP matches any packet whose destination is in that set, regardless of direction, so it also drops outgoing packets from containers to those IPs.

You’re absolutely right on both points:

With network_mode: “container:XYZ”, there is no “between-containers” network hop. Both containers share the same network namespace (same interfaces, IPs, routing, conntrack, and iptables). There’s nothing to firewall “between” them at L3/L2—the kernel sees only one stack.

Alright I will confess that I didn’t know this. This piece of info from chatgpt changes what you want to do from “significantly non trivial” to “basically impossible”. This means that containers do not have seperate ip addresses/networking for you to isolate from each other, they all share a single network namespace. You would have to isolate traffic based on other factors, like the process ID or user ID, which are not really inherently tied to the container.

As a bonus:

Docker’s ICC setting historically controls inter-container comms on bridge networks (default bridge or a user-defined bridge with enable_icc=). It doesn’t universally control every mode, and it won’t help when two containers share a netns.

Useful for understanding terminology I guess, but there is a class of these problems these tools really struggle to solve. I like to assign problems like this to people and then they will often attempt to use chatgpt at first, but then they will get frustrated and quickly realize chatgpt is not an alternative for using your brain.

mirror.gcr.io is google’s public mirror of dockerhub.

https://moonpiedumplings.github.io/blog/docker-registry/

There are a few apps that I think fit this use case really well.

Languagetool is a spelling and grammer checker that has a server client model. Libreoffice now has built in languagetool integration, where it can acess a server of your choosing. I make it access the server I run locally, since archlinux packages languagetool.

Another is stirling-pdf. This is a really good pdf manipulation program that people like, that comes as a server with a web interface.

I think I have also seen socket access in Nginx Proxy Manager in some example now. I don’t really know the advantages other than that you are able to use the container names for your proxy hosts instead of IP and port

I don’t think you need socket access for this? This is what I did: https://stackoverflow.com/questions/31149501/how-to-reach-docker-containers-by-name-instead-of-ip-address#35691865

I’ve seen three cases where the docker socket gets exposed to the container (perhaps there are more but I haven’t seen any?):

1. Watchtower, which does auto updates and/or notifies people

2. Nextcloud AIO, which uses a management container that controls the docker socket to deploy the rest of the stuff nextcloud wants.

3. Traefik, which reads the docker socket to automatically reverse proxy services.

Nextcloud does the AIO, because Nextcloud is a complex service, but it grows to be very complex if you want more features or performance. The AIO handles deploying all the tertiary services for you, but something like this is how you would do it yourself: https://github.com/pimylifeup/compose/blob/main/nextcloud/signed/compose.yaml . Also, that example docker compose does not include other services, like collabara office, which is the google docs/sheets/slides alternative, a web based office.

Compare this to the kubernetes deployment, which yes, may look intimidating at first. But actually, many of the complexities that the docker deploy of nextcloud has are automated away. Enabling the Collabara office is just collabara.enabled: true in the configuration of it. Tertiary services like Redis or the database, are included in the Kubernetes package as well. Instead of configuring the containers itself, it lets you configure the database parameters via yaml, and other nice things.

For case 3, Kubernetes has a feature called an “Ingress”, which is essentially a standardized configuration for a reverse proxy that you can either separate out, or one is provided as part of the packages. For example, the nextcloud kubernetes package I linked above, has a way to handle ingresses in the config.

Kubernetes handles these things pretty well, and it’s part of why I switched. I do auto upgrade, but I only auto upgrade my services, within the supported stable release, which is compatible for auto upgrades and won’t break anything. This enables me to get automatic security updates for a period of time, before having to do a manual and potentially breaking upgrade.

TLDR: You are asking questions that Kubernetes has answers to.

Try the yaml language server by red hat, it comes with a docker compose validator.

But in general, off the top of my head, dashes = list. No dashes is a dictionary.

So this is a list:

thing:
    - 1
    - 2

And this is a dictionary:

dict:
    key1: value1
    key2: value2

And then when they can be combined into a list of dictionaries.

listofdicts:
    - key1dict1: value1dict1
    - key1dict2: value1dict2
      key2dict2: value2dict2

And then abother thing to note is that yaml wilL convert things into a string. So if you have ports 8080:80, this will be converted into a string, which is a clue that this is a string in a list, rather than a dictionary.

deleted by creator

There is uksmd for ram dedupe.

Many helm charts, like authentik or forgejo integrate bitnami helmcharts for their databases. So that’s why this is concerning to me,

But, I was planning to switch to operators like cloudnativepostgres for my databases instead and disable the builtin bitnami images. When using the builtin bitnami images, automatic migration between major releases is not supported, you have to do it yourself manually and that dissapointed me.

I’m on my phone rn and can’t write a longer post. This comment is to remind me to write an essay later. I’ve been using authentik heavily for my cybersecurity club and have a LOT of thoughts about it.

The tldr about authentik’s risk of enshittification is that authentik follows a pattern I call “supportware”. It’s when extremely (intentionally/accidentally) complex software (intentionally/accidentally) lacks edge cases in their docs,because you are supposed to pay for support.

I think this is a sustainable business model, and I think keycloak has some similar patterns (and other Red Hat software).

The tldr about authentik itself is that it has a lot of features, but not all of them are relevant to your usecase, or worth the complexity. I picked up authentik for invites (which afaik are rare, also official docs about setting up invites were wrong, see supportware), but invites may not something you care about.

Anyway. Longer essay/rant later. Despite my problems, I still think authentik is the best for my usecase (cybersecurity club), and other options I’ve looked at like zitadel (seems to be more developer focused),or ldap + sso service (no invites afaik) are less than the best option.

Sidenote: Microsoft entra is offers similar features to what I want from authentik, but I wanted to self host everything.

Straying away from utilities, games are always fun to host. I got started with self hosting by hosting a minecraft server, but there are plenty of options.

So instead you decided to go with Canonical’s snap and it’s proprietary backend, a non standard deployment tool that was forced on the community.

Do you avoid all containers because they weren’t the standard way of deploying software for “decades” as well? (I know people that actually do do that though). And many of my issues about developers and vendoring, which I have mentioned in the other thread I linked earlier, apply to containers as well.

In fact, they also apply to snap as well, or even custom packages distributed by the developer. Arch packages are little more than shell scripts, Deb packages have pre/post hooks which run arbitrary bash or python code, rpm is similar. These “hooks” are almost always used for things like installing. It’s hypocritical to be against curl | bash but be for solutions like any form of packages distributed by the developers themselves, because all of the issues and problems with curl | bash apply to any form of non-distro distributed packages — including snaps.

You are are willing to criticize bash for not immediately knowing what it does to your machine, and I recognize those problems, but guess what snap is doing under the hood to install software: A bash script. Did you read that bash script before installing the microk8s snap? Did you read the 10s of others in the repo’s used for doing tertiary tasks that the snap installer also calls?

# Try to symlink /var/lib/calico so that the Calico CNI plugin picks up the mtu configuration.

The bash script used for installation doesn’t seem to be sandboxed, either, and it runs as root. I struggle to see any difference between this and a generic bash script used to install software.

Although, almost all package managers have commonly used pre/during/post install hooks, except for Nix/Guix, so it’s not really a valid criticism to put say, Deb on a pedestal, while dogging on other package managers for using arbitrary bash (also python gets used) hooks.

But back on topic, in addition to this, you can’t even verify that the bash script in the repo is the one you’re getting. Because the snap backend is proprietary. Snap is literally a bash installer, but worse in every way.

Except k3s does not provide a deb, a flatpak, or a rpm.

Canonical’s snap use a proprietary backend, and comes at a risk of vendor lock in to their ecosystem.

The bash installer is fully open source.

You can make the bad decision of locking yourself into a closed ecosystem, but many sensible people recognize that snap is “of the devil” for a good reason.

I’ve tried snap, juju, and Canonical’s suite. They were uniquely frustrating and I’m not interested in interacting with them again.

The future of installing system components like k3s on generic distros is probably systemd sysexts, which are extension images that can be overlayed onto a base system. It’s designed for immutable distros, but it can be used on any standard enough distro.

There is a k3s sysext, but it’s still in the “bakery”. Plus sysext isn’t in stable release distros anyways.

Until it’s out and stable, I’ll stick to the one time bash script to install Suse k3s.

I find this comparison unfair becuase k3s is a much more batteries included distro than the others, coming with an ingress controller (traefik) and a few other services not in talos or k0s.

But I do think Talos will end up the lighest overall because Talos is not just a k8s distro, but also a extremely stripped down linux distro. They don’t use systemd to start k8s, they have their own tiny init system.

It should be noted that Sidero Labs is the creator of Talos Linux, which another commenter pointed out.

https://wiki.archlinux.org/title/Arch_is_the_best

I think that distributing general software via curl | sh is pretty bad for all the reasons that curl sh is bad and frustrating.

But I do make an exception for “platforms” and package managers. The question I ask myself is: “Does this software enable me to install more software from a variety of programming languages?”

If the answer to that question is yes, which is is for k3s, then I think it’s an acceptable exception. curl | sh is okay for bootstrapping things like Nix on non Nix systems, because then you get a package manager to install various versions of tools that would normally try to get you to install themselves with curl | bash but then you can use Nix instead.

K3s is pretty similar, because Kubernetes is a whole platform, with it’s own package manager (helm), and applications you can install. It’s especially difficult to get the latest versions of Kubernetes on stable release distros, as they don’t package it at all, so getting it from the developers is kinda the only way to get it installed.

Relevant discussion on another thread: https://programming.dev/post/33626778/18025432

One of my frustrations that I express in the linked discussion is that it’s “developers” who are making bash scripts to install. But k3s is not just developers, it’s made by Suse who has their own distro, OpenSuse, using OpenSuse tooling. It’s “packagers” making k3s and it’s install script, and that’s another reason why I find it more acceptable.