I think they understate the importance of accepting OCI and Dockerfile semantics as a path to an external "run one of these" and having it actually emerge as a jail based outcome.
I get saying "we don't need these additional layers/abstractions" but what it ignores is me saying "I want to run this code, and what I have is a suite of Docker based behaviour and I want a low friction path to use that Docker compose method, to get where I want"
They also haven't yet addressed how things re-scale sideways. Pods, and scaling is why people wind up behind traefik or caddy, fronting a service. It's not because the service lies in RFC1918 (how I wish they had written kubernetes to V6 native) it's because the service is being delivered by multiple discrete runtime states "inside" and scales horizontally.
It's a different operating system. You can't point at a dockerfile, say "port this please from linux-such-and-such to FreeBSD" and expect it to work every time. There are nuances even with linux-compat.
Contrary to popular belief load-balance/scaleout is orthogonal to containers (and k8s is only one of the ways to go about it), so obviously it's not discussed in an article about containers.
Very often you can, or could, because the software is portable (e.g. Node or Python or Postgres), and / or platform-independent (e.g. written in JS, Python, bash, etc).
In my practice it was completely normal to build things inside a container to be deployed on Linux using the same sources and basically the same package names and versions as used on a developer macOS machine (which is BSD-like enough down below).
> macOS machine (which is BSD-like enough down below)
That's like saying an Ubuntu .deb will work on Gentoo because it's all Linux anyway. It's not that simple. There is dependencies and there are differences in the packages, package managers and surrounding system for a reason. It's not 1:1.
Perhaps the naming scheme happened to line up for the packages you where using, but this should be considered not assumed.
It would be nice if there was some sort of translator that could handle "most common cases". I think it would improve the usability of Jails.
Perhaps that would require someone to keep a list of packages mapping certain packages between operating systems.
Something like "apt install python3-serial" -> "pkg install py311-pyserial" may suffice.
For anyone that would use something like that, you should implement a prototype, publish it and perhaps someone else will build upon what you started!
The article is written by the people who created jail.run [0]
Ten years in with Docker and Linux containers I felt something was very wrong so I looked at how Solaris and FreeBSD were doing it and I saw the light (too).
I would agree with them: bringing the Dockerfile format to jails doesn't make any sense, unless you just want to attract curious Linux users.
Dockerfiles are useful and familiar but are also an abomination.
What we need is solid way to do configuration management. I guess this is what they are trying to do with their own configuration system [1] but I am not sold on it yet.
If said system also targets Linux then that would be cool. Otherwise we just get another new way to do something on one OS and doesn’t at all help bridge the gap.
Daemonless is a bit the linuxserver.io of FreeBSD to me. It is a collection of standardized images.
Behind each image you have a Containerfile [0] and a compose.yaml [1]
And inside the Containerfile for Gitea you simply have a RUN pkg install -y gitea [2], so basically here it is the good old FreeBSD Gitea port [3]
I guess this is really a Linux/Docker users friendly wrapper on the 30 years old FreeBSD ecosystem?
When I came from Linux to FreeBSD, Gitea was one of the first service I ran in a jail to learn how it all works. What I was happy to discover is I just need to do pkg -j mygiteajail install gitea to install Gitea in the jail with all the rc scripts, etc.
The jail abstraction was just one option (-y) in pkg ! This is really the beauty of the all integrated FreeBSD. So to be honest the Daemonless have, in some case at least, made everything more complicated in my view...
I'll bite: how do we take advantage of ZFS layering if not via the docker-style layering?
I find dockerfile layering to be unsatisfying because step 5 might depend on step 2 but not 3 or 4... the linearisation of a DAG makes them harder to maintain and harder to cache cleanly (with us also having monster single-line CMDs all in the main of image results).
one thing that bit me with LXC: anything that needs its own kernel module won't work. jails have the same limitation — shared host kernel. ran into this trying to run a VPN server (needs DKMS for a custom wireguard fork) in an LXC container — module can't load, period. ended up on a full KVM VM.
The main drawback I saw on jails is that they are FreeBSD. The owner doesn’t mention, and I have not researched it, but can you run any Linux distribution in a FreeBSD jail?
I would like to explore the interoperability/compatibility limits of LXC and OCI support in FreeBSD 15. Both with FreeBSD as an OCI container and Linux OCI containers within FreeBSD.
30 comments
I get saying "we don't need these additional layers/abstractions" but what it ignores is me saying "I want to run this code, and what I have is a suite of Docker based behaviour and I want a low friction path to use that Docker compose method, to get where I want"
They also haven't yet addressed how things re-scale sideways. Pods, and scaling is why people wind up behind traefik or caddy, fronting a service. It's not because the service lies in RFC1918 (how I wish they had written kubernetes to V6 native) it's because the service is being delivered by multiple discrete runtime states "inside" and scales horizontally.
Contrary to popular belief load-balance/scaleout is orthogonal to containers (and k8s is only one of the ways to go about it), so obviously it's not discussed in an article about containers.
In my practice it was completely normal to build things inside a container to be deployed on Linux using the same sources and basically the same package names and versions as used on a developer macOS machine (which is BSD-like enough down below).
> macOS machine (which is BSD-like enough down below)
That's like saying an Ubuntu .deb will work on Gentoo because it's all Linux anyway. It's not that simple. There is dependencies and there are differences in the packages, package managers and surrounding system for a reason. It's not 1:1. Perhaps the naming scheme happened to line up for the packages you where using, but this should be considered not assumed.
It would be nice if there was some sort of translator that could handle "most common cases". I think it would improve the usability of Jails. Perhaps that would require someone to keep a list of packages mapping certain packages between operating systems.
Something like "apt install python3-serial" -> "pkg install py311-pyserial" may suffice.
For anyone that would use something like that, you should implement a prototype, publish it and perhaps someone else will build upon what you started!
> It's not that simple.
It would tremendously benefit almost everyone if it were.
> There is dependencies and there are differences in the packages, package managers and surrounding system for a reason.
Yeah, the NIH syndrome. And sometimes, of course, there are decent technical reasons as well.
Ten years in with Docker and Linux containers I felt something was very wrong so I looked at how Solaris and FreeBSD were doing it and I saw the light (too).
I would agree with them: bringing the Dockerfile format to jails doesn't make any sense, unless you just want to attract curious Linux users.
Dockerfiles are useful and familiar but are also an abomination.
What we need is solid way to do configuration management. I guess this is what they are trying to do with their own configuration system [1] but I am not sold on it yet.
Anyway those people are doing some good work !
- [0] https://jail.run/
- [1] https://jail.run/reference/jail-config/
[0] https://daemonless.io/
Behind each image you have a Containerfile [0] and a compose.yaml [1] And inside the Containerfile for Gitea you simply have a
RUN pkg install -y gitea[2], so basically here it is the good old FreeBSD Gitea port [3]I guess this is really a Linux/Docker users friendly wrapper on the 30 years old FreeBSD ecosystem?
When I came from Linux to FreeBSD, Gitea was one of the first service I ran in a jail to learn how it all works. What I was happy to discover is I just need to do
pkg -j mygiteajail install giteato install Gitea in the jail with all the rc scripts, etc.The jail abstraction was just one option (
-y) in pkg ! This is really the beauty of the all integrated FreeBSD. So to be honest the Daemonless have, in some case at least, made everything more complicated in my view...- [0] https://github.com/daemonless/gitea/blob/main/Containerfile....
- [1] https://github.com/daemonless/gitea/blob/main/compose.yaml
- [2] https://github.com/daemonless/gitea/blob/9bb9151d31ae6574e5f...
- [3] https://cgit.freebsd.org/ports/tree/www/gitea
I find dockerfile layering to be unsatisfying because step 5 might depend on step 2 but not 3 or 4... the linearisation of a DAG makes them harder to maintain and harder to cache cleanly (with us also having monster single-line CMDs all in the main of image results).
So is there a better way that people are using?
With IBM and Unisys big iron predating ever further in the historic evolution, outside UNIX based operating systems.
Perhaps formalisation lends values, which deploy in the analysis of field research.
Code 11
Identification of C++ develeopers
Compilation of C++ code into Python
Recursive .py scripts that are sector-enframings of the general "neural" framework deployment.
[1]:https://malus.sh/blog.html