Have there been any good efforts into getting rid of the language or providing an alternative? My effort into switching and the biggest complaint I've read is that the idea behind NixOS and the ecosystem is genius but one of the biggest draw backs is writing everything in Nix.
I spend a weekend every so often defining the core of what I want next time I upgrade, but just find it so annoying I'm sure I won't use anything I've written until there's a major change in the ecosystem.
Putting aside the poor typing (the lack of proper typing is a shame, so valid criticism), I actually really like the language - it's genuinely a great DSL for the particular problems it's supposed to handle.
It does take a bit of use for it to click, though. A lot of it has to do not with Nixlang itself but about learning nixpkgs' idioms.
This is inherently a Hard Problem™, since completions may require evaluating arbitrary derivations (e.g. building a custom Linux kernel).
For "what symbols are available", the nil LSP implementation[1] works for anything in scope that doesn't require evaluation. It also includes completions for the stdlib and NixOS options (in certain contexts).
Another LSP implementation is nixd[2], which is trying to tackle the problem of evaluations for completion.
Like (iirc) systemd-resolved has `enabled` which is false by default but then gets silently turned on if you use systemd-networkd. How are you supposed to figure that out without reading the source?
But I think this also stems from the fact that the default state of nixos is "a general purpose linux system" and so instead of just starting at 0 and adding the things you need, you have to mix adding and removing things which IMO makes things much more complicated (except maybe for newbies to linux who don't know what's necessary for a running system).
More to your point, though: I think a lot is possible. Although nix is very dynamic, it is also, for all intents and purposes, side effect free. I've had this idea that a sufficiently advanced IDE should be able to evaluate your nix code and tell you exactly what the possible values (not just types, but value!) are for any particular variable.
> I've had this idea that a sufficiently advanced IDE should be able to evaluate your nix code and tell you exactly what the possible values (not just types, but value!)
Similarly to the REPL, I'm often using `nix-instantiate --eval -E 'somethingsomething'` so it should definitely be possible.
Way too much sugar/"idioms", which makes it hard for someone new to the language to figure out what a given piece of code is actually doing. Confusing use of semicolons for what almost every other language uses commas or newlines (or nothing) for. It's the same feeling as writing bash, and needing to always look up again exactly what the syntax is and where the semicolons go.
It seems inevitable to me that some of the design choices around immutability and isolation are going to result in a larger server image (both on disk and in memory) than if you are prepared to forgo those things. For most people that tradeoff is probably worth it but if you want something to run in an embedded server or with a very low disk footprint it's probably not right for you.
Around 20 years ago people who wanted to do this[2] used to make tiny immutable redhat servers by remounting /usr and a few other things read-only after boot so it's certainly doable but it's a lot more of a pain than what nix does and there is no process isolation and no rollback etc when things go wrong.
[1] ...or generally in fact but that's a matter of opinion and I know people feel differently about this.
I find the nix language to be quite pleasant. There are some syntax quirks and types would be nice, but in general the “json with functions” vibe is imo great and a very nice fit for the domain. Lots of other modern config languages (e.g. dhall, jsonnet) have ended up in this part of the design space too.
With that said tweag has been working on a kind of nix 2.0 / nix with types for a while with the aim (I think) of being able to use it in nixpkgs: https://github.com/tweag/nickel
I also quite like nixlang for config tasks - in theory! In practice its really annoying. I think the main problem is the interpreter and the bad error messages / bad debuggability.
Part of that just comes from lazy evaluation, which makes debugging a lot harder in general (you feel this in Haskell...), but also just from nix not being a big popular language that gets lots of polish, and being completely dynamically typed.
There is Guix, which replaces the Nix language with Scheme, but which has some limitations related to a smaller user base, e.g. a smaller package collection.
Replacing the language requires duplicating all the work that went into Nix, to reach parity, so it is not easy.
The biggest limitation IMO it's that they are HPC-centric, not caring the desktop, which is the was to allow more people discovering a distro. Also the lack of a proper zfs and lvm/mdraid/luks support it's a big showstopper.
> Replacing the language requires duplicating all the work that went into Nix, to reach parity, so it is not easy.
That seems like a design flaw in Nix, there's no reason the data model should be so tightly coupled to the scripting implementation that you can't reuse packages written in a different language.
There is no technical barrier against doing that. But much of the power and flexibility in nixpkgs arises from the nix language, not the data model (which is comparatively simple).
Using a different language to depend on packages derived from .nix would be very much akin to depending on a docker image whose Dockerfile you can not inspect.
> Using a different language to depend on packages derived from .nix would be very much akin to depending on a docker image whose Dockerfile you can not inspect.
Speaking of Docker images and Dockerfiles, that's actually a real-world example of how you can achieve this kind of effect without relying on a specific language. Ironically, you can use Nix to build Docker images; there's a bunch of other alternative builders (e.g. Kaniko, Buildah); you can also just stitch together some files&metadata into tarballs, and then 'docker import' it.
Nix or Guix are of course much more powerful and expressive than Docker images, but there's always a cost to complexity.
It seems to be an issue with testing and debugging, rather than the language itself. The same issue would also be present if you could switch to any other language for configuration.
Nickel lang is such an effort. Id say the syntax is a mix of json and lua and aims for a non-touring complete program. It is still a bit early but it looks promising
No, Nickel is Turing-complete. That's been one of the characteristics intended to distinguish it from most other configuration languages from the start.
I don't know any other OS where you could even go on such a trip so easily! Figuring out why things are where and being able to disable them like this is pretty cool.
I was really surprised you were able to replace systemdMinimal with systemd in dbus though.
I thought it was there to break the cyclic dependency between systemd and dbus
Depends on what you consider "easy"! Nix itself has a pretty high barrier to entry.
Personally I believe systems that start simple (e.g. Alpine) are easier to mess with. Plus you don't have to give up all benefits of declarative configuration; for example apk has a single file (/etc/apk/world) that defines the exact package set that needs to remain installed. You can edit it and run "apk fix", much like you can edit /etc/nixos/configuration.nix and rerun "nixos-rebuild switch". It's not as powerful as Nixos, but power (and complexity) always has a price.
He didn't mention that nix uses a lot of ram. If your server is tiny and doesn't have swap enabled, running nix command s will make it unresponsive. Are there any nixos alternative that allow you to do system wide configuration from a single source similar to configuration.nix?
As others said, I've moved away from doing nix builds on servers and into a less wasteful (if you're running multiple servers) approach of building once, deploying the bits into a cache, and making the servers fetch them from a cache. I've been slowly working on my own tool to make this workflow easier (nixless-agent, you can find an early version on GitHub).
If you have a different machine with more ram and compute, you can use 'nixos-rebuild --target-host=<server> switch'. That does all the nix building on the local machine, and then just copies binaries and activates the built configuration on the remote machine.
You need not perform your nix evaluation on the same device you are targeting. You can nix copy a system closure to your target and activate it, and there are a number of tools in the nix ecosystem to make this easier.
Granted, if you local machine is low on RAM, or isn't Linux, then you will be in trouble.
I haven't tried it personnally but Guix is similar with config in guile scheme. Have a look at the documentation [1][2].
Caveat: it is a gnu project so no proprietary stuff like firmwares and drivers included out of the box (but there is a community guix nonfree project available [3]). I believe that isn't a problem for virtual machine servers anyway.
Not similar to nix, but you can look into Yocto. You can use it to generate an OS. It is, much more involved than using nix, but suitable for low memory environments.
Well, "loved" is not the word I would use. It's an improvement over the chaos before, where everyone used some other bespoke tool to build their BSPs. Yocto at least is some kind of standard now, but not a particularly good one (granted, it's a really difficult problem to solve). I don't know anyone actually enjoying working with Yocto.
I just finished moving all of my servers off NixOS and I'm finally breathing a sigh of relief.
Deterministic systems are a cool idea, but we're just not there yet. The headaches and pain involved in maintaining these systems and warping the software to obey are too great.
Everything in NixOS works, until it doesn't. And when it doesn't, woe be unto you.
What did you replace it with? I've been been using NixOS for a (non-critical) server which replaces another server that was "managed" by Ansible. Now that's it running nicely on NixOS, and while it's far from perfect, I'm really struggling to see how I would manage it in any other way. The experience is so much nicer now.
And yes, I have put a lot of blood sweat and tears into making things work in nix/NixOS. The thing that keeps me invested is once I get something working, it is far easier to keep it working. If nixpkgs updates break my things, I'm one git bisect away from figuring out what happened.
I use Proxmox on top of Debian. Surprisingly, I'm back to Bash scripts to set things up (because Ansible wouldn't allow me to orchestrate between a host and a guest - so if I have to change container config during setup, I'm back to scripts again, AND the complications of Ansible).
I basically build up Proxmox container templates, and then build upon those similar to how Docker does it (I don't use Docker because they don't allow you to specify your MAC address, so you can't control them from a separate LAN-based DHCP server - instead you have to map a bunch of ports on your host and then configure all external clients to match... so dumb).
I've basically gone full circle at this point:
- Docker
- LXD with Bash scripts
- LXD with a ton of Python
- NixOS
- Proxmox with Ansible
- Proxmox with Bash scripts (albeit much simpler and flatter than last time)
Everything is containerized and has its own IP address on the physical LAN, the templates can be regenerated with a simple script, important data is mapped to a host directory (/home/data/my-container, which gets backed up), and destroying and rebuilding an instance container is a cinch.
One really nice feature of this setup is that I can tear down and rebuild a template, launch a test container from that, copy the instance data in /home/data to the new container, make sure it works with the new stuff, and then launch it for real.
Now it doesn't matter what technology (container or VM) I use. Everything is a completely separate machine as far as the LAN is concerned, which greatly simplifies things.
Everything, from host to software to containers & VMs is built "deterministically" (i.e. deterministically enough) from the scripts. Rebuilding the whole thing (server and all) from scratch takes about an hour and a half. I just use the same set of scripts on all of my servers to make management easier. Hosts have minimal software and configuration, and guests do all the real work. Migrating is an rsync /home/data away.
As someone who uses Proxmox without bash scripts and is scared of the day I have to re-work out all the config files I (often randomly through trial and error) changed:
Do you have any tips on how to get started?
Do you simply make the change and then paste the commands needed into a script?
Also I assume you have a script to set up the (Proxmox) host machine?
I also have quite a few Proxmox specific things that I had to change (e.g. GPU pass-through) which seems to break your "Now it doesn't matter what technology (container or VM) I use" advantage.
Proxmox really needs to get their cloud-init story together. If they supported cloud-init for LXC a lot of automation and setup issues would just go away…
I agree. The title made me expect a low-quality, clickbait article. The actual article provides a lot of educational value to users of Nix and those interested in trying to build the smallest Linux distribution that can run Nix. Despite using Nix (the package manager) for so long, I haven't considered what it would take to get this. The article answers the question perfectly.
> I was trying to bring NixOS to a bare minimum, which is an exercise similar to building containers with the bare minimum required for the software in the container to run. I think this is a worthy endeavour. I think we have all the tools in regular non-docker, non-kubernetes linux to get to a similar outcome, except we won’t need docker or kubernetes or whatever in this new land, thus removing quite a bunch of complexity from the systems we build.
> But doing it on top of NixOS currently feels like a bad path to take.
The author of this blog post might be interested in playing with not-os, another, much smaller OS built with Nix: https://github.com/cleverca22/not-os
Thanks! I have to admit that I've had the itch to build my own NixOS-inspired system more than once, and I haven't done that because I just don't have time to dedicate to this among all the other projects I'm working on. I wasn't aware of not-os before, but I'll definitely dig into the code!
How "stable" are these kinds of slimming tweaks though? In a rolling release setup, aren't you going to have extraneous packages semi-regularly drag in all of Python or Perl accidentally? The setup might break and you'd not even notice?
I find the premise of a carefully re-compilable/re-creatable system very appealing, but not having a stable LTS style release rather incongruous. It takes a huge effort to get all the pieces working together - and if it's rolling and the sand is shifting/breaking underneath you it feels you never reach a meaningful stable system. Sure you can recreate your well tested working configuration, but the configurations is effectively immediately out of date and unmaintained once any packages are updated
I think this is why they effectively only target x64. I'm not a "distro guy" so maybe I'm missing something. It seems it'd be sensible to just 1-to-1 copy Ubuntu LTS package versions (+ patches) and build a NixOS "stable" version that can be patched and updated as people find issues
> I think this is why they effectively only target x64
Nix and Nixpkgs is the best in class when it comes to cross platform & cross architecture support. It has good support for x86_64 / aarch64 /macOS / Linux. Getting Musl or static variants of existing packages just work for many packages. There's even some work on BSD / Windows support. Cross compiling is far easier to setup compared to other package managers. If anything, other projects should be copying what Nix is doing.
Almost all of the changes flip an official setting. Those stay around for a long time and get a proper deprecation notice when they go away, so you won't be surprised. Replacing systemd-minimal with the full version may potentially cause some edge case issues, but it's the same package with more features enabled, so I wouldn't really expect any.
Nothing will break when the package gets updated as long as you keep to your specific release - backported changes are backwards compatible.
- a language friendly as Haskell, so while fit for purpose definitively it's not well digested by most, also by various longtime NixOS users;
- an unclear direction, there are countless of "side projects" and no clear path, most are not even indexed in a wiki page so you just discover by accident interacting with someone else or after a search;
- a terrible documentation probably due to the lack of a clear direction stated above.
The biggest "mean install" is true, but it's not that much impacting in the real world, NixOS real purpose is AVOIDING containers in designing an infra, not being wrapped by them or wrapping them and true x86 zero-overhead virtualization does not exists. So far only IBM Power Systems with AiX seems to have something nearly-zero-overhead built-in in the (big)iron.
IMVO that's the main point: most people, NixOS devs included, fails to see a world different than the current one. A possible answer could be keeping up the evolution of zfs and mirror some IllumOS features so we can have light paravirtualization thanks to zones on zfs clones. But as per NixOS most people fails to see a different storage than the most common today, a relict model from the '80s (does anyone remember the infamous "zfs is a rampant layer violation" phrase?). A damn real modern system should be: a SINGLE application, yes, the OS as a framework, development environment who produce a running system live out of itself. A coupled package-manager/installer/storage, because those are effectively a unique thing so we do not need a network of symlinks or containers, we have a storage behind that simply expose needed software pieces together, also a system who manage in-memory stuff the same way. Zfs was the first step in this direction, with boot environments, clones, zones glued by the Image Package System, lacking the language for a proper system integration, unfortunately almost nobody have taken care of that. NixOS and Guix System offer another piece, the language to integrate package management, installers but they lack the storage integration to generate a unique new system model.
Rediscover IllumOS (OpenSolaris) would bridge the gap providing all needed piece to start a new kind of distro and infra management for a FLOSS world where there is no need of monsters to deploy simple infra and those simple tools could scale at monster level, killing the commercial IT model of the giants and given the humanity the desktop model, the "pioneering internet" of interconnected personal system model a new start.
The lack of independent universities, big labs, is probably the root cause but as always good things tend to happen anyway sooner or later, it's the interim the bad part.
- didn't check if nixos uses it, but coreutil has a single binary mode (like busybox, a single binary is built and symlinks or hardlinks are used to provide all the commands); that might save some space
- instead of trying to strip down the system, maybe go the other way around: only include the command you need with its closure?
closure computation is done in a few places (apparmor profile or systemd.confinement come to mind) and it should be possible to just copy whatever your server binary needs, your kernel (since microVM and not container), and run the binary directly as init (maybe with a simple wrapper that hardcodes network or whatsnot)
> - didn't check if nixos uses it, but coreutil has a single binary mode (like busybox, a single binary is built and symlinks or hardlinks are used to provide all the commands); that might save some space
It does. If you have coreutils from Nixpkgs installed you can check this with
basename "$(realpath "$(which ls)")"
If you see 'ls', you're looking at separate binaries. If you see 'coreutils', it's a symlink to a singular 'coreutils' binary.
I'm not sure about just the one closure since the result needs to boot, but more generally that looks reasonable if your use case fits. There's a couple efforts, IIRC, mostly centered around router firmware where "real" nixos doesn't fit, ex. https://gti.telent.net/dan/liminix
2 is very interesting to me. I'd love to see an example of a "scratch" vm build, in a similar spirit to scratch docker builds. Assume I have all my software and kernel built, what is the minimal filesystem that will boot? How much easier can it get when I am only targeting a VM?
After the blog post the system was around 600mb, but what would be a reasonable image size for a server? Also is this bare metal or in the cloud (for the worker machines)?
One technique employed by microvm.nix[0] is to mount the hosts /nix/store into the guest. This won't shrink the size of the system, but should allow it to be amortized across many different VMs.
I'm not sure how exploitable a read-only virtiofs share is, so this is perhaps not appropriate in some circumstances.
I was thinking of a similar approach, but mounting /nix/store from the host into the guest will only work if you have a single guest.
For multiple guests, you should rely instead on:
* A snapshot-able filesystem with the option to create clones (like ZFS). I think this is a great idea actually.
* Exporting /nix/store via NFS, so you can have multiple writers (but this creates some tight coupling in that accidentally deleting stuff there may disrupt all guests).
On borderline, thinking to learn nix. Few questions:
Can I define VM images with nix? Can those VM images be loaded into VirtualBox?
Also, possible to do similar to build AMIs or other cloud VMs such as Hetzner where my "Dockerfile" is a nix file which defines the system to be built and then it has everything in place once built including tools, libraries, configuration and such?
The ecosystem is in my experience very well fleshed out (7 yrs of use), as long as you don't require a knowledgebase/wiki/ up2date documentation, it's not been a issue for me since I could always fall back on Linux knowledge and just looking for how other distributions do x / how the thing itself is configured , and looking at how perhaps a existing nix module wraps that
Disk space is why one of my laptops doesn't run NixOS; 16GB soldered storage fit one generation passably, but it was clear that it wasn't gonna work (of course, that's with X and a browser; totally different use case). I vaguely recall that locales were shockingly large and I couldn't figure out how to trim them down. So that box runs Alpine now, which easily fits with no effort.
What is the security update policy with NixOS? I've always had the impression that I would need to upgrade every 6 months to keep getting critical security fixes, which is unappealing compared to e.g. Debian + Nix
I know I can revert easily if there are problems when upgrading, but that doesn't really apply if security fixes only land in the new branch
You get backported security fixes in release channels as well. Unless anything changed recently, there's no explicit guarantee around them, but the core packages typicaly land just as fast or faster then other distros. Keep in mind though that more esoteric software with a small number of users and auto-update disabled may lag a bit.
Maybe I'm missing something, but the only branch in github:nixos/nixpkgs I can see receiving fixes is the 24.05 branch getting fixes backported from unstable. The last commit I can see to the 23.11 branch is about 3 months ago
This would imply only 9 months of security patches before I would need to upgrade the server. That is of course a far less risky process with NixOS, so perhaps that is ok, but it is a lot more work than the 5 years you get (free) with Ubuntu/Debian
Taking the time to delve in to things like this can teach you more about the OS you run than many other things can. It's not too unlike making a crunchgen OS, or even just a read-only OS with overlays. It's sometimes quite enlightening to see what will work with much less than we assumed it needs.
Having just installed the entirety of NetBSD on an i386 system (a 200 MHz Pentium), I see it weighs in at around 1 gig. But that's everything including X11 with WM and the toolchain (gcc 10). That's not bad, but it's really amazing how much of that isn't necessary for running the OS on a server. Particularly where you might want tiny VM images
Have there been any good efforts into getting rid of the language or providing an alternative? My effort into switching and the biggest complaint I've read is that the idea behind NixOS and the ecosystem is genius but one of the biggest draw backs is writing everything in Nix.
I spend a weekend every so often defining the core of what I want next time I upgrade, but just find it so annoying I'm sure I won't use anything I've written until there's a major change in the ecosystem.
What's so bad about it?
Putting aside the poor typing (the lack of proper typing is a shame, so valid criticism), I actually really like the language - it's genuinely a great DSL for the particular problems it's supposed to handle.
It does take a bit of use for it to click, though. A lot of it has to do not with Nixlang itself but about learning nixpkgs' idioms.
It's not just poor typing. It's the lack of discoverability. "What are all my options here in this expression? What are the symbols available?"
I think a good IDE integration could solve this, but not sure how much is possible.
This is inherently a Hard Problem™, since completions may require evaluating arbitrary derivations (e.g. building a custom Linux kernel).
For "what symbols are available", the nil LSP implementation[1] works for anything in scope that doesn't require evaluation. It also includes completions for the stdlib and NixOS options (in certain contexts).
Another LSP implementation is nixd[2], which is trying to tackle the problem of evaluations for completion.
[1] https://github.com/oxalica/nil/
[2] https://github.com/nix-community/nixd
Like (iirc) systemd-resolved has `enabled` which is false by default but then gets silently turned on if you use systemd-networkd. How are you supposed to figure that out without reading the source?
But I think this also stems from the fact that the default state of nixos is "a general purpose linux system" and so instead of just starting at 0 and adding the things you need, you have to mix adding and removing things which IMO makes things much more complicated (except maybe for newbies to linux who don't know what's necessary for a running system).
The nix repl can be a very valuable tool in answering these questions.
That said, I strive to structure my nix source so that portions of it can easily be pasted into a repl. ReadTree goes a long way in that regard: https://github.com/tvlfyi/kit/tree/canon/readTree
More to your point, though: I think a lot is possible. Although nix is very dynamic, it is also, for all intents and purposes, side effect free. I've had this idea that a sufficiently advanced IDE should be able to evaluate your nix code and tell you exactly what the possible values (not just types, but value!) are for any particular variable.
> I've had this idea that a sufficiently advanced IDE should be able to evaluate your nix code and tell you exactly what the possible values (not just types, but value!)
Similarly to the REPL, I'm often using `nix-instantiate --eval -E 'somethingsomething'` so it should definitely be possible.
There is this: https://github.com/nix-community/nixd
It has jump to definition and autocomplete. Which is very nice.
It's not perfect. But it's pretty good
Way too much sugar/"idioms", which makes it hard for someone new to the language to figure out what a given piece of code is actually doing. Confusing use of semicolons for what almost every other language uses commas or newlines (or nothing) for. It's the same feeling as writing bash, and needing to always look up again exactly what the syntax is and where the semicolons go.
Here's all the cases for using a semicolon. (estimated 30 seconds read)
1. At the end of local variables
2. At the end of each attributes in an attribute set (a.k.a. dictionaries or key-value pairs) 3. with expressions 4. Assertions Now, you'll never be confused again.I don't think the language is the issue here[1].
It seems inevitable to me that some of the design choices around immutability and isolation are going to result in a larger server image (both on disk and in memory) than if you are prepared to forgo those things. For most people that tradeoff is probably worth it but if you want something to run in an embedded server or with a very low disk footprint it's probably not right for you.
Around 20 years ago people who wanted to do this[2] used to make tiny immutable redhat servers by remounting /usr and a few other things read-only after boot so it's certainly doable but it's a lot more of a pain than what nix does and there is no process isolation and no rollback etc when things go wrong.
[1] ...or generally in fact but that's a matter of opinion and I know people feel differently about this.
[2] me for one, but others also.
I find the nix language to be quite pleasant. There are some syntax quirks and types would be nice, but in general the “json with functions” vibe is imo great and a very nice fit for the domain. Lots of other modern config languages (e.g. dhall, jsonnet) have ended up in this part of the design space too.
With that said tweag has been working on a kind of nix 2.0 / nix with types for a while with the aim (I think) of being able to use it in nixpkgs: https://github.com/tweag/nickel
I also quite like nixlang for config tasks - in theory! In practice its really annoying. I think the main problem is the interpreter and the bad error messages / bad debuggability.
Part of that just comes from lazy evaluation, which makes debugging a lot harder in general (you feel this in Haskell...), but also just from nix not being a big popular language that gets lots of polish, and being completely dynamically typed.
There is Guix, which replaces the Nix language with Scheme, but which has some limitations related to a smaller user base, e.g. a smaller package collection.
Replacing the language requires duplicating all the work that went into Nix, to reach parity, so it is not easy.
The biggest limitation IMO it's that they are HPC-centric, not caring the desktop, which is the was to allow more people discovering a distro. Also the lack of a proper zfs and lvm/mdraid/luks support it's a big showstopper.
> Replacing the language requires duplicating all the work that went into Nix, to reach parity, so it is not easy.
That seems like a design flaw in Nix, there's no reason the data model should be so tightly coupled to the scripting implementation that you can't reuse packages written in a different language.
There is no technical barrier against doing that. But much of the power and flexibility in nixpkgs arises from the nix language, not the data model (which is comparatively simple).
For example, see zb: https://www.zombiezen.com/blog/2024/09/zb-early-stage-build-...
Using a different language to depend on packages derived from .nix would be very much akin to depending on a docker image whose Dockerfile you can not inspect.
> Using a different language to depend on packages derived from .nix would be very much akin to depending on a docker image whose Dockerfile you can not inspect.
Speaking of Docker images and Dockerfiles, that's actually a real-world example of how you can achieve this kind of effect without relying on a specific language. Ironically, you can use Nix to build Docker images; there's a bunch of other alternative builders (e.g. Kaniko, Buildah); you can also just stitch together some files&metadata into tarballs, and then 'docker import' it.
Nix or Guix are of course much more powerful and expressive than Docker images, but there's always a cost to complexity.
It seems to be an issue with testing and debugging, rather than the language itself. The same issue would also be present if you could switch to any other language for configuration.
Nickel lang is such an effort. Id say the syntax is a mix of json and lua and aims for a non-touring complete program. It is still a bit early but it looks promising
No, Nickel is Turing-complete. That's been one of the characteristics intended to distinguish it from most other configuration languages from the start.
See the 'RATIONALE' document: https://github.com/tweag/nickel/blob/378ece30b3e3c0ab488f659...
> Have there been any good efforts into getting rid of the language or providing an alternative?
Guix is conceptually similar to Nix but uses scheme.
I don't know any other OS where you could even go on such a trip so easily! Figuring out why things are where and being able to disable them like this is pretty cool.
I was really surprised you were able to replace systemdMinimal with systemd in dbus though.
I thought it was there to break the cyclic dependency between systemd and dbus
Depends on what you consider "easy"! Nix itself has a pretty high barrier to entry.
Personally I believe systems that start simple (e.g. Alpine) are easier to mess with. Plus you don't have to give up all benefits of declarative configuration; for example apk has a single file (/etc/apk/world) that defines the exact package set that needs to remain installed. You can edit it and run "apk fix", much like you can edit /etc/nixos/configuration.nix and rerun "nixos-rebuild switch". It's not as powerful as Nixos, but power (and complexity) always has a price.
He didn't mention that nix uses a lot of ram. If your server is tiny and doesn't have swap enabled, running nix command s will make it unresponsive. Are there any nixos alternative that allow you to do system wide configuration from a single source similar to configuration.nix?
(I'm the post author)
As others said, I've moved away from doing nix builds on servers and into a less wasteful (if you're running multiple servers) approach of building once, deploying the bits into a cache, and making the servers fetch them from a cache. I've been slowly working on my own tool to make this workflow easier (nixless-agent, you can find an early version on GitHub).
If you have a different machine with more ram and compute, you can use 'nixos-rebuild --target-host=<server> switch'. That does all the nix building on the local machine, and then just copies binaries and activates the built configuration on the remote machine.
I don't run nixos on my PC
you don't need to have nixos to have nixos-rebuild installed. Just nix
You need not perform your nix evaluation on the same device you are targeting. You can nix copy a system closure to your target and activate it, and there are a number of tools in the nix ecosystem to make this easier.
Granted, if you local machine is low on RAM, or isn't Linux, then you will be in trouble.
I haven't tried it personnally but Guix is similar with config in guile scheme. Have a look at the documentation [1][2].
Caveat: it is a gnu project so no proprietary stuff like firmwares and drivers included out of the box (but there is a community guix nonfree project available [3]). I believe that isn't a problem for virtual machine servers anyway.
[1] guix cookbook: https://guix.gnu.org/cookbook/en/html_node/index.html#Top
[3] guix manual: https://guix.gnu.org/manual/en/html_node/index.html#Top
[3] https://github.com/guix-users/guix-nonfree
The widely used nonfree guided channel is https://gitlab.com/nonguix/nonguix, not the one you linked.
Not similar to nix, but you can look into Yocto. You can use it to generate an OS. It is, much more involved than using nix, but suitable for low memory environments.
The problem with nix is that compiling a system uses lots of memory, but when deployed there's little overhead.
Like you would with Yocto, I just build my systems on a proper host then remotely deploy them.
Yocto is also used by Lenovo to build their software appliances like XClarity Administrator.
Also it’s very famous and loved in embedded software circles.
Yocto is also hated for its unnessesary complexity in embedded software circles.
Well, "loved" is not the word I would use. It's an improvement over the chaos before, where everyone used some other bespoke tool to build their BSPs. Yocto at least is some kind of standard now, but not a particularly good one (granted, it's a really difficult problem to solve). I don't know anyone actually enjoying working with Yocto.
author does not use nix on his nixos and removed it, so no ram issue.
author does not activate new config on host machine, but deploys new host machines as needed.
he evals on build/dev/ci machine only.
It won't necessarily do that if you deploy to your server from a remote machine over ssh
guix?
Doesn't it use nixos underneath?
guix does not use nixos underneath, it's a completely independent implementation (of the same underlying concepts)
guix is built on top of the nix daemon
It does have the nix daemon in-tree: https://git.savannah.gnu.org/cgit/guix.git/tree/nix
I just finished moving all of my servers off NixOS and I'm finally breathing a sigh of relief.
Deterministic systems are a cool idea, but we're just not there yet. The headaches and pain involved in maintaining these systems and warping the software to obey are too great.
Everything in NixOS works, until it doesn't. And when it doesn't, woe be unto you.
What did you replace it with? I've been been using NixOS for a (non-critical) server which replaces another server that was "managed" by Ansible. Now that's it running nicely on NixOS, and while it's far from perfect, I'm really struggling to see how I would manage it in any other way. The experience is so much nicer now.
I'll echo this. I can't imaging going back.
And yes, I have put a lot of blood sweat and tears into making things work in nix/NixOS. The thing that keeps me invested is once I get something working, it is far easier to keep it working. If nixpkgs updates break my things, I'm one git bisect away from figuring out what happened.
I use Proxmox on top of Debian. Surprisingly, I'm back to Bash scripts to set things up (because Ansible wouldn't allow me to orchestrate between a host and a guest - so if I have to change container config during setup, I'm back to scripts again, AND the complications of Ansible).
I basically build up Proxmox container templates, and then build upon those similar to how Docker does it (I don't use Docker because they don't allow you to specify your MAC address, so you can't control them from a separate LAN-based DHCP server - instead you have to map a bunch of ports on your host and then configure all external clients to match... so dumb).
I've basically gone full circle at this point:
- Docker
- LXD with Bash scripts
- LXD with a ton of Python
- NixOS
- Proxmox with Ansible
- Proxmox with Bash scripts (albeit much simpler and flatter than last time)
Everything is containerized and has its own IP address on the physical LAN, the templates can be regenerated with a simple script, important data is mapped to a host directory (/home/data/my-container, which gets backed up), and destroying and rebuilding an instance container is a cinch.
One really nice feature of this setup is that I can tear down and rebuild a template, launch a test container from that, copy the instance data in /home/data to the new container, make sure it works with the new stuff, and then launch it for real.
Now it doesn't matter what technology (container or VM) I use. Everything is a completely separate machine as far as the LAN is concerned, which greatly simplifies things.
Everything, from host to software to containers & VMs is built "deterministically" (i.e. deterministically enough) from the scripts. Rebuilding the whole thing (server and all) from scratch takes about an hour and a half. I just use the same set of scripts on all of my servers to make management easier. Hosts have minimal software and configuration, and guests do all the real work. Migrating is an rsync /home/data away.
As someone who uses Proxmox without bash scripts and is scared of the day I have to re-work out all the config files I (often randomly through trial and error) changed:
Do you have any tips on how to get started? Do you simply make the change and then paste the commands needed into a script?
Also I assume you have a script to set up the (Proxmox) host machine?
I also have quite a few Proxmox specific things that I had to change (e.g. GPU pass-through) which seems to break your "Now it doesn't matter what technology (container or VM) I use" advantage.
Edit: I've made the repo public: https://github.com/kstenerud/proxmox-containers
Here's one of my host setup files (run this immediately after installing Proxmox from the iso):
The host is kept VERY simple.Here's my basic Ubuntu container template:
Here's my Plex template (with GPU passthrough): Here's the Plex instance script (starts a Plex instance going from the Plex template):Maybe chuck that on GitHub/gist (or some other online repo) so people can potentially suggest stuff in an easy way? :)
Hmm OK. I already had it on github so I'll just make it public:
https://github.com/kstenerud/proxmox-containers
Proxmox really needs to get their cloud-init story together. If they supported cloud-init for LXC a lot of automation and setup issues would just go away…
What didn't work in your NixOS servers?
The post is much more interesting than the title suggests. Great deep dive into slimming a NixOS system.
I agree. The title made me expect a low-quality, clickbait article. The actual article provides a lot of educational value to users of Nix and those interested in trying to build the smallest Linux distribution that can run Nix. Despite using Nix (the package manager) for so long, I haven't considered what it would take to get this. The article answers the question perfectly.
> I was trying to bring NixOS to a bare minimum, which is an exercise similar to building containers with the bare minimum required for the software in the container to run. I think this is a worthy endeavour. I think we have all the tools in regular non-docker, non-kubernetes linux to get to a similar outcome, except we won’t need docker or kubernetes or whatever in this new land, thus removing quite a bunch of complexity from the systems we build.
> But doing it on top of NixOS currently feels like a bad path to take.
The author of this blog post might be interested in playing with not-os, another, much smaller OS built with Nix: https://github.com/cleverca22/not-os
Could be a decent source of inspiration!
(I'm the post author)
Thanks! I have to admit that I've had the itch to build my own NixOS-inspired system more than once, and I haven't done that because I just don't have time to dedicate to this among all the other projects I'm working on. I wasn't aware of not-os before, but I'll definitely dig into the code!
How "stable" are these kinds of slimming tweaks though? In a rolling release setup, aren't you going to have extraneous packages semi-regularly drag in all of Python or Perl accidentally? The setup might break and you'd not even notice?
I find the premise of a carefully re-compilable/re-creatable system very appealing, but not having a stable LTS style release rather incongruous. It takes a huge effort to get all the pieces working together - and if it's rolling and the sand is shifting/breaking underneath you it feels you never reach a meaningful stable system. Sure you can recreate your well tested working configuration, but the configurations is effectively immediately out of date and unmaintained once any packages are updated
I think this is why they effectively only target x64. I'm not a "distro guy" so maybe I'm missing something. It seems it'd be sensible to just 1-to-1 copy Ubuntu LTS package versions (+ patches) and build a NixOS "stable" version that can be patched and updated as people find issues
> I think this is why they effectively only target x64
Nix and Nixpkgs is the best in class when it comes to cross platform & cross architecture support. It has good support for x86_64 / aarch64 /macOS / Linux. Getting Musl or static variants of existing packages just work for many packages. There's even some work on BSD / Windows support. Cross compiling is far easier to setup compared to other package managers. If anything, other projects should be copying what Nix is doing.
Almost all of the changes flip an official setting. Those stay around for a long time and get a proper deprecation notice when they go away, so you won't be surprised. Replacing systemd-minimal with the full version may potentially cause some edge case issues, but it's the same package with more features enabled, so I wouldn't really expect any.
Nothing will break when the package gets updated as long as you keep to your specific release - backported changes are backwards compatible.
NixOS has stable releases built every 6 months.
NixOS IME have few issues:
- a language friendly as Haskell, so while fit for purpose definitively it's not well digested by most, also by various longtime NixOS users;
- an unclear direction, there are countless of "side projects" and no clear path, most are not even indexed in a wiki page so you just discover by accident interacting with someone else or after a search;
- a terrible documentation probably due to the lack of a clear direction stated above.
The biggest "mean install" is true, but it's not that much impacting in the real world, NixOS real purpose is AVOIDING containers in designing an infra, not being wrapped by them or wrapping them and true x86 zero-overhead virtualization does not exists. So far only IBM Power Systems with AiX seems to have something nearly-zero-overhead built-in in the (big)iron.
IMVO that's the main point: most people, NixOS devs included, fails to see a world different than the current one. A possible answer could be keeping up the evolution of zfs and mirror some IllumOS features so we can have light paravirtualization thanks to zones on zfs clones. But as per NixOS most people fails to see a different storage than the most common today, a relict model from the '80s (does anyone remember the infamous "zfs is a rampant layer violation" phrase?). A damn real modern system should be: a SINGLE application, yes, the OS as a framework, development environment who produce a running system live out of itself. A coupled package-manager/installer/storage, because those are effectively a unique thing so we do not need a network of symlinks or containers, we have a storage behind that simply expose needed software pieces together, also a system who manage in-memory stuff the same way. Zfs was the first step in this direction, with boot environments, clones, zones glued by the Image Package System, lacking the language for a proper system integration, unfortunately almost nobody have taken care of that. NixOS and Guix System offer another piece, the language to integrate package management, installers but they lack the storage integration to generate a unique new system model.
Rediscover IllumOS (OpenSolaris) would bridge the gap providing all needed piece to start a new kind of distro and infra management for a FLOSS world where there is no need of monsters to deploy simple infra and those simple tools could scale at monster level, killing the commercial IT model of the giants and given the humanity the desktop model, the "pioneering internet" of interconnected personal system model a new start.
The lack of independent universities, big labs, is probably the root cause but as always good things tend to happen anyway sooner or later, it's the interim the bad part.
Couple of things spring to mind:
- didn't check if nixos uses it, but coreutil has a single binary mode (like busybox, a single binary is built and symlinks or hardlinks are used to provide all the commands); that might save some space
- instead of trying to strip down the system, maybe go the other way around: only include the command you need with its closure? closure computation is done in a few places (apparmor profile or systemd.confinement come to mind) and it should be possible to just copy whatever your server binary needs, your kernel (since microVM and not container), and run the binary directly as init (maybe with a simple wrapper that hardcodes network or whatsnot)
good luck!
> - didn't check if nixos uses it, but coreutil has a single binary mode (like busybox, a single binary is built and symlinks or hardlinks are used to provide all the commands); that might save some space
It does. If you have coreutils from Nixpkgs installed you can check this with
If you see 'ls', you're looking at separate binaries. If you see 'coreutils', it's a symlink to a singular 'coreutils' binary.> go the other way around
I'm not sure about just the one closure since the result needs to boot, but more generally that looks reasonable if your use case fits. There's a couple efforts, IIRC, mostly centered around router firmware where "real" nixos doesn't fit, ex. https://gti.telent.net/dan/liminix
2 is very interesting to me. I'd love to see an example of a "scratch" vm build, in a similar spirit to scratch docker builds. Assume I have all my software and kernel built, what is the minimal filesystem that will boot? How much easier can it get when I am only targeting a VM?
This might be something you're interested in: https://appfs.rkeene.org/web/dir?ci=trunk&name=build/qemu
It's for running AppFS inside a VM on QEMU. It uses a statically linked Tcl (which AppFS is written in) to bring the system up.
After the blog post the system was around 600mb, but what would be a reasonable image size for a server? Also is this bare metal or in the cloud (for the worker machines)?
One technique employed by microvm.nix[0] is to mount the hosts /nix/store into the guest. This won't shrink the size of the system, but should allow it to be amortized across many different VMs.
I'm not sure how exploitable a read-only virtiofs share is, so this is perhaps not appropriate in some circumstances.
[0]: https://github.com/astro/microvm.nix
I was thinking of a similar approach, but mounting /nix/store from the host into the guest will only work if you have a single guest.
For multiple guests, you should rely instead on: * A snapshot-able filesystem with the option to create clones (like ZFS). I think this is a great idea actually. * Exporting /nix/store via NFS, so you can have multiple writers (but this creates some tight coupling in that accidentally deleting stuff there may disrupt all guests).
Have you taken a look at https://github.com/nix-community/srvos ? It’s a collection of profiles for servers by the Numtide folks.
On borderline, thinking to learn nix. Few questions:
Can I define VM images with nix? Can those VM images be loaded into VirtualBox? Also, possible to do similar to build AMIs or other cloud VMs such as Hetzner where my "Dockerfile" is a nix file which defines the system to be built and then it has everything in place once built including tools, libraries, configuration and such?
Thoughts?
EDIT: Typos
Yep for sure! We even have a neat upload-ami cli tool:
https://github.com/NixOS/amis
yes to all.
https://github.com/nix-community/nixos-generators
And for cloud/dedicated servers perhaps https://github.com/nix-community/nixos-anywhere
(There's also a terraform module ) And for state changes https://github.com/serokell/deploy-rs Or Colmena /nixops/ and x other
For secret handling perhaps https://github.com/ryantm/agenix/ + https://github.com/oddlama/agenix-rekey
The ecosystem is in my experience very well fleshed out (7 yrs of use), as long as you don't require a knowledgebase/wiki/ up2date documentation, it's not been a issue for me since I could always fall back on Linux knowledge and just looking for how other distributions do x / how the thing itself is configured , and looking at how perhaps a existing nix module wraps that
Disk space is why one of my laptops doesn't run NixOS; 16GB soldered storage fit one generation passably, but it was clear that it wasn't gonna work (of course, that's with X and a browser; totally different use case). I vaguely recall that locales were shockingly large and I couldn't figure out how to trim them down. So that box runs Alpine now, which easily fits with no effort.
Have you tried putting /nix on an filesystem with compression? I've experienced >50% in storage savings with Bcachefs + zstd and dedupe
I too use Alpine. I like how you start small and add what you need. The option of having a read-only root FS is convenient on flaky SBCs.
What is the security update policy with NixOS? I've always had the impression that I would need to upgrade every 6 months to keep getting critical security fixes, which is unappealing compared to e.g. Debian + Nix
I know I can revert easily if there are problems when upgrading, but that doesn't really apply if security fixes only land in the new branch
You get backported security fixes in release channels as well. Unless anything changed recently, there's no explicit guarantee around them, but the core packages typicaly land just as fast or faster then other distros. Keep in mind though that more esoteric software with a small number of users and auto-update disabled may lag a bit.
Maybe I'm missing something, but the only branch in github:nixos/nixpkgs I can see receiving fixes is the 24.05 branch getting fixes backported from unstable. The last commit I can see to the 23.11 branch is about 3 months ago
This would imply only 9 months of security patches before I would need to upgrade the server. That is of course a far less risky process with NixOS, so perhaps that is ok, but it is a lot more work than the 5 years you get (free) with Ubuntu/Debian
https://github.com/NixOS/nixpkgs/branches/active
Good stuff! From the title I thought this would be about sysadmin woes, but this is interesting too :)
Another little addendum: you can trivially create bootable custom NixOS installer images with whatever configuration you want pre-applied[0].
[0] https://nix.dev/tutorials/nixos/building-bootable-iso-image
You can just use them as immutable live systems too, with the unfortunate requirement that it'll call your system an installer: https://github.com/NixOS/nixpkgs/issues/338738
Taking the time to delve in to things like this can teach you more about the OS you run than many other things can. It's not too unlike making a crunchgen OS, or even just a read-only OS with overlays. It's sometimes quite enlightening to see what will work with much less than we assumed it needs.
Having just installed the entirety of NetBSD on an i386 system (a 200 MHz Pentium), I see it weighs in at around 1 gig. But that's everything including X11 with WM and the toolchain (gcc 10). That's not bad, but it's really amazing how much of that isn't necessary for running the OS on a server. Particularly where you might want tiny VM images
This is highly relevant for me, will probably end up using some of the same tricks they've done to trim the size of some of my micro VMs.
Could you run ’nix store optimise’ before you remove ’nix’?
This reminds me a little bit of trying to slim Docker images.
Any OS without stable releases and a security team is not a server OS.
https://status.nixos.org/ https://nixos.org/community/teams/security/
Welcome to NixOS, then! :)
NixOS does have "stable" releases every 6 months, but they don't have long term support.
Then you only have Ubuntu, Fedora, Red Hat/CentOS/Alma/Rocky and Suse/OpenSuse.