Linux

I encourage you to go to town with whatever crazy setup you come up.

I just want to note that the reboot-to-update mechanism also has its positive sides, as ancient as it may seem (we do not succumb to windows level backwardness, because that fails to reap the benefits despite requiring so many reboots). Namely, you get atomic updates, hence the name "fedora atomic" for example. That means you have no transient periods where your OS is running in an inconsistent state. Like when you update a traditional distro, the new files/libraries/binaries/kernel-modules do not match anymore what is in RAM, including the currently running kernel. That leads to stuff like the nvidia driver / cuda not working until reboot, running applications failing to load a library they need now etc.. The vast majority of times this is no huge problem, but in theory the only way of maintaining a system with it never running in basically undefined state is with atomic udpates.

My semi-immutable OS is based around a Debian installation where every application is installed in a separate Docker container.

When you launch the application, it volume mounts an appropriate directory that contains only the data related to that application.

Chrome for example launches with a single subdirectory inside ~/Downloads, so each instance can only see its own directory.

I can also test compilation of random repositories inside a container, without affecting the underlying OS.

The OS itself has only got a minimal Debian and Docker installed.

Been using it for several years. I can't recall when I last rebooted it.

Most immutable systems I’ve seen require a reboot in order to apply system changes.

It depends on the type of change you're applying and which specific immutable/atomic/composable distro we're talking about. In the case of GuixSD and NixOS, this is simply (mostly) false. For the others, if you're going to change the base-system/image, then a reboot is required. Though, it's important to mention that often enough systemctl soft-reboot suffices. And even then, some of them (like Fedora Atomic) offer functionality (e.g. through --apply-live) to apply the changes directly to the working system. And we're not even talking about the latest innovations (like systemd-sysext that might be able to provide a revolutionary solution to this problem altogether.

Also, even immutable OS I’ve seen wants you to opt-in to a rollback. Rarely do I see the full effects of installing a package or altering a config immediately. By the time I notice an issue maybe it’s too late to rollback to before the change or maybe I’ve done a few other things since and I don’t want to rollback everything. I would much prefer to make “rolling forward” or persisting changes to be a very conscious process.

You seem to have false notions on how this works on atomic distros, or at least it seems your understanding only applies to a small subset of them. E.g. responsible use of GuixSD and/or NixOS completely circumvents this from the get-go. Even Fedora Atomic offers (relatively) sane management of rollbacks.

I started messing with BTRFS and I think I’ve come up with a process that will get me what I want, no matter the distro. Please poke holes in my idea. So I think I can use BTRFS to hold data for the rootfs in three different subvolumes (at minimum): root-A, root-B, root-Z. root-Z is my golden image and it represents what I want root to look like after reboot. root-A and root-B are the active and passive instances of rootfs, but which one is active will flip-flop after every reboot. So if I boot with A, B gets replaced with the contents of Z. In the meantime I can do whatever I want with A. Not sure how I’ll update Z (chroot or “promote” the active subvol to be Z) but without an update every reboot is an automatic rollback.

I don't understand how this solves your (conceived) issue with rollbacks. Furthermore, I also wonder what kinda changes you're applying to your system on a daily basis. Like, do you really feel the need to poke at /usr every day? Aren't your needs regarding installing software already (mostly) solved through AppImage/Nix/Flatpak/Snap/Brew etc?

BTRFS snapshots are relatively cheap. I have about 10 or 20 or so on my systems. Snapper manages that for me.

Also I'd say that you only want to roll back for catastrophic errors like the system not booting. If you discover after 8 upgrades that a package that was updated on the first upgrade has some critical bug you wouldn't roll back everything else to fix the issue. You would just downgrade that single package to a known good version.

OpenSUSE microOS/ microOS Desktop (Aeon, Kalpa) does this.

They use a complete "changes go to the next system" thing also using BTRFS.

But they dont use OSTree so the system is fundamentally flawed.

Advantages of ostree are

• complete transparency over package changes rpm-ostree db diff
• complete transparency over /etc changes (the upstream is in /usr/etc and can be reset, see here
• the OS is always based on a complete upstream remote, your local system does not matter at all. You can rebase, reset etc without being dependent on anything on the local OS.

Example: I could rebase from Fedora OSTree to CentOS OSTree. They are working on bootc images, which are bootable OCI images and in theory only one step away from uBlue-like distribution.

If you do anything relying on local package management like OpenSUSE does, you can snapshot between changes but still mess up.

So I would always base off OSTree.

What I dont get though is the reliance on reboots and images. OSTree works on all filesystems and doesnt need images, it is simply like a Git repo.

So what I would change is, to enable random local changes with a flag --direct and simply apply the changes live. I mean, that is what DNF and all the distros do too.

Only if you need a kernel upgrade you do stuff with a reboot. Version upgrades are also WAY better than the unstable mess on standard Fedora or other distros.

So track everything with OSTree, allow resets, rebases etc, but dont force all the image stuff. This is the reason why rpm-ostree takes so long and is so inefficient compared no DNF.

Just using OSTree you could only install RPMs, use a nonwheel user, SELinux confined users and have a secure and slim system.

I dont know if I miss something here. Android is rootless but the base OS is still immutable and uses A/B root, so writing only happens to the inactive partition. I dont know if immutability is some core security feature.

Rpm-ostree is really good as an allrounder, but I think a bit overkill. It does support installing packages live, but this does the same action afaik and just swaps the OS image without a reboot.

VanillaOS uses something similar with ABroot + OCI images, except for the flip flopping at every reboot.