We're working in this space with Project Bluefin: https://github.com/projectbluefin/dakota
Both approaches are indeed competitive, but you can also leverage both to achieve the same thing. We're experimenting with a pure ddi Bluefin, a buildstream/GNOMEOS one that spits out a bootc image, as well as a Bluefin that is just a systemd-sysext on top of GNOME OS. Chef's choice!
There will be many ways to slice this problem -- my opinion is that in the end it will be how you design the infrastructure to make these and not the artifacts themselves.
We already have CentOS/Fedora builds alongside these, long term we'll see which ones end up being the most efficient. Buildstream is a tool which people should look at in this space too: https://buildstream.build/index.html
> Contributing members include people from Ubuntu Core, Debian, GNOME OS, Fedora CoreOS, Endless OS, Arch Linux, SUSE, Flatcar, systemd, image-builder/osbuild, mkosi, tpm2-software, System Transparency, buildstream, BTRFS, bootc, composefs, (rpm-)ostree, Microsoft, Amazon, and Meta.
Note systemd, (rpm-)ostree and bootc.
My understanding is that uapi is another initiative but not completely separated from bootc and ostree. Maybe complementary.
I'm struggling to see how sysupdate is really equivalent to bootc or ostree though. Sysupdate is just the sw-update tool from Yocto that's been around for 10+ years with a little more syntax sugar, which itself was just a common shared implementation of what all embedded systems had been rolling-thier-own of for almost 20 years before that. It says it requires an A/B/.../N partitioning scheme, which is exactly what bootc/ostree/etc is designing to avoid.
If you don't use the whole disks update thing from sysupdate, then instead you're just talking about a transactional package manager that is still in its infancy and hasn't addressed the many gotcha and corner cases that the dozens of more mature package managers have. It's not actually "transactional" in the sense of undo for example, it's "transactional" only in that you won't get partial install, which hasn't been a problem with any existing package managers for almost 40 years. All thier listed things you can list together for a "transaction" association are either things that are already linked via existing package maager packages, or are only useful for embedded systems.
I'm not saying sysupdate isn't useful, upper end of embedded design is pushing into the space where systemd is standard now so it could be useful for those devices, but it's not really equivalent at all to bootc/ostree, and doesn't really have amt applicability outside initial system imaging from a live disk, or embedded devices.
These are words but they don't make sense.
Context re:distros mentioned:
GNOME OS: https://os.gnome.org/ KDE Linux: https://kde.org/linux/
They are also considering moving to buildstream and join gnome.
To elaborate, I will separate linux operating systems into three categories:
1. traditional mutable package management (debian, fedora, arch, etc) - Bad model for obvious reasons, I won't get into this.
2. immutable images (embedded deployments, android, chromeos) - A build system (yocto, nix, buildroot, buildstream) procudes a disk image. System boots the disk image and handles updates using A/B root partitions with `systemd-sysupdate` or similar.
3. store-based atomic (nixos, ublue) - The system keeps a mutable store of hashed objects (/nix/store, /var/lib/containers/storage, /ostree/repo) and boots a specific system generation. Updates add new objects to the store, which must be automatically cleaned, and create a new systemd generation to boot into.
In the case of categories 2 and 3, a build system of some kind is used to produce the image or packages that are turned into the new system generation.
The bootc project, which falls into category 3, attempts to use the standardized and highly adopted OCI image format (layered filesystem changes stored in a content addressed store) as a medium for distributing linux systems. The major limitation here is that these systems are very complicated to build and extend. While it may not seem that way compared to nixos if you have prior experience writing dockerfiles, configuring your system with imperative statements that build on previous state is _really_ tedious. For example, you can't just re-use work with multiple `FROM` statements in the same layer, so you instead need to copy files between images. This is incredibly jank, look at [the docs](https://blue-build.org/how-to/minimal-setup/) for bluebuild's module system. Additionally, for a motivated user to change their system internals, they need to make the jump to hosting it with CI and pulling the images.
As jcastro mentioned elsewhere in this thread, there is work in the ublue project, which focuses on bootc images, to instead build their systems from source using buildstream, the same way GNOME OS works. The idea is that this is no longer a "distribution" and doesn't have "packages" anymore, since the entire system is built from source into an image that can be updated to atomically. While this model is simpler and way less jank than assembling your OS in a dockerfile, the model of separating packaging from distribution makes things harder for users for no benefit over the alternative.
To elaborate: Buildstream is a build system created by gnome developers that works as follows:
- Various element are defined in yaml files with their dependency and build steps
- Buildstream forms a tree of build elements to evaluate
- Each element is evaluated in a sandbox with access to only its dependencies and the output is placed in a content-addressed store for caching
If you have experience with nix, you may recognize that this is almost the same way nix works, with the difference that buildstream dependencies are mounted to regular directories (e.g. /usr) in the build sandbox instead of directly dealt with as /nix/store objects, and that buildstream is much harder to extend. A nix build results in artifacts that live in the nix store and are symlinked to other paths in the nix store, while a buildstream build results in artifacts that are compatible with a more traditional directory structure.
The idea of GNOME OS being "distroless" and not having packages is misleading, as it does have packages. Only, the packages exist exclusively during the build process. In order for a user to modify their system, they must add a rule to the buildstream definitions their system is built from and rebuild the entire thing from source to generate a new system image, which is unnecessarily burdensome. This is because the content-addressed artifact store (buildstream's cache) that exists when building has no relationship to the content-addressed artifact store on the deployed system (ostree or oci storage). This is pointless indirection with no clear benefit. By not separating building from distribution, nixos (a project using the nix dsl to build a linux system) achieves the same benefits of this model without any of the drawbacks. Users can modify how their system works, use caches, add their own packages, and share/integrate these modifications freely without building their entire os again from source every time.
To put this power in perspective, adding a nginx server to your system on nixos amounts to adding to your nix configuration:
services.nginx.enable = true;
and then rebuilding, switching to the new system generation.
Imagine how painful it would be to do this via a dockerfile (based off a different image) or buildstream definitions...
Finally, I'd like to clarify that nixos is not perfect. There are many areas of that need improvement (documentation, evaluation speed, evaluation caching, ifd, error messages, etc). However, I believe nixos is a fundamentally better and simpler model than the one being pushed by a lot of the linux desktop ecosystem at the moment. I believe a lot of the work on infrastructure like bootc and buildstream would be better focused on nix/nixos, or at least would benefit from learning from them.
first off i can start nginx locally i dont need to change a fancy config in a bad DSL(and yes its bad actually horrible),but if i want it in an image i add one line in my docker file RUN systemctl enable nginx and rebuild thats it does it take longer to rebuild WHO CARES CI takes care of it
May I rephrase that?
bootc and OStree are both Cthulhoid nightmare horrors that only exist because of corporate politics, but the leading edge...
"quite a bit easier to extend" sounds good to me, but the "easier" here refers to the internal system implementation details? I am an end user, not a Linux distro system architect, and I care more about the user experience. I will be interested in test driving a general purpose OS based on this technology, whenever that happens in the future. Since Red Hat is involved in the UAPI project, perhaps Fedora Atomic Desktop will migrate to this technology in the future?
I really like the NAS use case because I can build the ZFS kmods for that specific version of Fedora CoreOS in CI/CD. If there's any compatibility failure, then my NAS doesn't get an update and I get the CI/CD failed email. No downtime because of some kernel incompatibility.
For the laptop though, I feel like there's a better way that I haven't found. Some way not to require CI/CD, to build the next image and switch to it all locally. I haven't gone down that path yet, but it looks kinda like that Option 2 the author described. Maybe it's really just that easy.
I've really been enjoying this space.
OCI roughly assumes that layers will be laid out in some logical way, and that a given host will see opportunities to reuse across different instances, but with bootc, there will only ever be one instance.
OCI also assumes that individual layers are small enough that it is always worth pulling and unpacking a layer instead of some kind of authentication delta, which is great for a k8s cluster in a center, but not great for devices out on the edge, where you might want this kind of pseudo-immutable system even more.
I really want some standardized way for a manifest in OCI to say that "this content is also available in other format X here".
rpm-ostree worked fine for me
It works quite well for our edge devices. It’s tightly integrated with Toradex hardware, but not limited to it.
It may seems litte a niche, but it has strong potential for long‑term supported edge products. Any additional experiences to share?
So same OS, users, packages, flatpaks etc. And a mostly synced home dir too.
Is NixOS the only viable way to do this? I don't like the path mangling that Nix introduces.
It seems like an immutable distro customized via a Containerfile could work too? Except rebooting/reimagine for every change sounds tedious as hell.
This does cause some interesting tension in the immutability though. /etc in particular is really a mix of things that a sysadmin should really only be setting, and things a regular user may set indirectly. This usage has grown organically over time with the tools involved in the implementation, so it's not at all consistent which are which. The immutable system solutions recognize this by usually handling the whole /etc folder the same way package managers handle package installs that include /etc file: by doing a 3-way merge between the old provided files, the new provided files, and the current existing files to see if the existing are unchanged from the old provided and can just be directly replaced by the new provided or if a merge conflict needs resolving. Additionally, a separate copy of /etc is maintained associated with each available bootable system version so when you roll back you get the old /etc files you had before. Though this does introduce a system-unique variation since you now have new /etc being affected by the state of /etc when it was forked.
If you want all your home folder and system config to be identical, nix or guix really are your primary way to go, that extra lockdown of the user and system config is exactly what most people don't want for usability reasons.
I personally use nix home-manager on top of Aurora DX from Universal Blue. I have my nix home-manager config setup to manage only the things I want to be locked down in my home config, and to provide some extra tools that are easier to manage/supply via Nix than a system package manager (where I would need to do a whole system update to get the new version). My IDE for example is installed on a specific version via Nix, but I don't have Nix manage the settings of it so I can separately tweak as needed without need a home-manager rebuild.
EDIT: typo
You can do this today with Aurora, Bazzite, Bluefin, and other bootc systems. The system updates by default are weekly and require a reboot but when you move most of the stuff into the userspace most of that stuff updates independently anyway.
Luckily UniversalBlue makes it incredibly easy, they have a template repo you can use that has all the GitHub action setup included to auto-bills on every change, and directions for how to set it up. It took me about 10 minutes
Next up, backups stored as layers in the same OCI registries.
I am not, however, sure ostree is going to be the final image format. Last time I looked work was in progress to replace that.
https://github.com/bootc-dev/bootc/issues/1190
There's a GitHub org that builds bootc-ready images for non-Red Hat family distributions using this backend.
With OSTree my team can deploy a commit over an existing ones to a field of 60k embedded Linux devices. Similarly bootc is a great alternative for deploying images for dedicated or virtualized servers.
The only reason Red Hat needed to invent this very complex mechanism was because RH does not officially have a COW-snapshot capable filesystem in its enterprise distro.
A filesystem with snapshots makes software installation transactional. You take a snapshot, install some software, and if it doesn't work right, you can revert to the snapshot. (With very slightly more flexible snapshots, you can limit the snapshot to just some part of the directory tree, but this is not essential; it merely permits more flexibility.)
In other words, you are a long way toward what in database language is called ACID:
https://en.wikipedia.org/wiki/ACID
Atomicity, consistency, isolation, durability. It makes your software inastallation transactional: an update either happens completely (A), you can check it is valid (C) and works (I), or it can be totally reverted, and the system restored to the earlier state (D).
That's a good thing. It means you can safely automate software deployment knowing that if it goes wrong you have an Undo mechanism. Databases got this 50+ years ago; in the 21st century it's making its way to FOSS OSes.
Do this in the filesystem and it's easy. SUSE's implementation is so simple, it's basically a bunch of shell scripts, and it can be turned on and off. You can run an immutable OS, reboot for updates, and if you need, disable it, go in and fix the system, and then turn it back on again.
This is because SUSE leans very heavily on Btrfs and that is the critical weakness -- Btrfs is only half finished and is not robust.
But RH removed Btrfs from RHEL and Btrfs was the only GPL COW filesystem, so core infrastructure in the distro means no COW on RH. Oracle Linux has Btrfs -- the FS was developed at Oracle, after all -- and so does Alma.
(Yes I know, Fedora put it back, but the key thing is, it only uses Btrfs only for compression so that Flatpak looks less horrendously inefficient. Fedora doesn't use snapshots.)
With no COW FS, RH had to invent a way to do transactional updates without filesystem support. Result, OStree. Git, but for binaries.
And yes, everyone developing FOSS uses Git, but almost nobody understands Git:
You know that if there's an Xkcd about it, it must be true.
Embedding something you don't understand in your OS design is a VERY BAD PLAN.
With OStree your FS is a virtual one, it's not real, it's synthesized on the fly from a local repository. The real FS is hidden and can't be hand-edited or anything. It generates the OS filesystem tree on the fly, you see. OS-tree.
Use it just for GUI apps, that's Flatpak.
Use it for the whole OS, that's OStree. It is so mind-shreddingly complicated that you can't do package management any more, you can't touch the underlying FS. So you need a whole new set of layers on top: virtual directories on top of the main virtual directory, and some bits with extra pseudo-filesystems layered on top of that to make some bits read-write.
It's like the scene in the Wasp Factory where under the skull plate it's just writhing maggots. I recall in horror and revulsion when I see it.
So it's deeply bizarre to read blog posts praising all the cool stuff you can do with it.
A legitimate point, but I feel that there must be better, cleaner, simpler, more elegant ways of doing this.
I mean, Nix does this and it's clean in its way, but the price is, a filesystem layout that is not navigable or maintainable by humans. IMHO and that of many people, that's a price too high, but it shows that there are alternative routes.
Yawn. You again. Hater!1!! ;-> Mine is still running, after all these (two) years abusing it. Even with installing(not really, much more complicated) some Windoze into a subvolume of it by means of WinBtrfs. I can either boot that Windows natively, or from the same install into a VM, by means of what once were called bootmanager profiles. Automagically chosen. Without having to reserve space for NTFS, VFAT(excluding UEFI/boot), or anything else. Its just a fucking folder, like any other.
> But RH removed Btrfs from RHEL and Btrfs was the only GPL COW filesystem
There is NILFS(2) which is GPL. Lacks some features, but that could have been worked around, or implemented?
> I recall in horror and revulsion when I see it.
You have total recall of recoiling in horror? What a trip... ;-)
That aside, it's all cargo-cult anyway, because of this thing called FHS. Which seems crazy, because in a world of open-source, which can be compiled almost any way one could wish for, why not just use slashpkg, Gobolinux or similar stuff?
Yep, me again.
You know that line from the James Bond books?
“Once is happenstance. Twice is coincidence. Three times is enemy action.”
I saw Btrfs die so many times, I know what to suspect.
The official internal guidance was: give it all your disk space. Make the volumes so big that they will never fill up.
Well, if I could trust it, I might, but I know I can't trust it, so I want my data safe somewhere else. And the company wouldn't provide me with 2 or 3 disk drives so `/home` could live on another physical volume.
And once you've watched your primary work OS self-destruct again, well, if you're me, you reinstall with dual-boot so you have something else to fall back on if it happens again. As it did happen again. Repeatedly.
> You have total recall of recoiling in horror? What a trip... ;-)
OK, my bad. Brain/finger error.
> why not just use slashpkg, Gobolinux or similar stuff?
Slashpkg means DJB's thing?
https://cr.yp.to/slashpackage.html
That sounds fantastic. Thanks. I had never heard of this before.
Re Gobo -- yes please. I love that little distro.
Eh, you don't typically have a lock mechanism for the filesystem equivalent to that of a database.
Who's to say something like this doesn't happen:
- snapshot fs
- op/system adjust firewall rules
- "you" install updates
- you rollback
- firewall rules is now missing patches
I never said it did!
But if you have a snapshotting FS underneath, transactional software maintenance becomes an order or two of magnitude easier to achieve.
The underlying philosophies of Unix are "keep it in files" and "keep it simple". That's why it didn't even have a file-hiding mechanism -- the dot-file thing was an accental, emergent property.
Keep it simple, keep it visible, keep it human-readable and human-fixable.
Because the more complex you make it, the more likely it is to go wrong, and some poor sap is going to have to fix it. Do not get in their way. Instead, think about them, allow for that, and help keep their life easy.