Has anyone used Nix or something similar?
I think cargo, rustup, nvm etc make things much easier in terms of getting build environments standardised, but this looks like another level.
I guess the biggest limitation with Nix is that it appears to be Linux only. Any thoughts?
Yes, I’ve tried nix and guix, mainly out of interest of deterministic builds for safe network (see this topic). So far neither have really seemed like ‘it’ to me. You’ll see from the link that bitcoin uses or is interested in using guix, I think they mainly rely on gitian at the moment)
I’ve tried chroot and firejail but neither really stuck for me, I’ve tried virtualbox and qubes-os but both were a little heavy handed (I really liked qubes but the overhead is still a bit much to swallow), I’ve not yet tried containers/docker though. All the benefits I was looking for from these things could be had by simply ‘knowing my tools’, and I’ve been able to run everything locally without conflict. All these extra tools didn’t justify their presence, but that’s obviously just a personal preference and I can see how they’re all still fantastic tools.
My main feeling on reproducible builds is that content-addressable libs packages and dependencies is going to be pretty handy to have. Nix has been using the content-addressable feature of IPFS (article) so it seems natural that this will also be utilized on Safe Network too.
Thanks @mav, I remember your interest in reproducible builds. Later I read a Twitter thread by a guy who was arguing that everything people say they want reproducible builds for is invalid. Have you seen that? I read the thread and think he’d written longer blogs about it, and it seemed convincing although I didn’t study it.
When I read about Nix I remembered all the time I spent trying to get one build environment or another working over the last few years and thought that it could be a good way to help others avoid that pain.
I hadn’t thought about this, very interesting, thanks for that. I suspect there’s a lot of things Safe CAFS will be good for that I haven’t anticipated!
No I haven’t, if you have a link to it I’d be keen to read.
Yeah I feel you on this, although much less so lately (I’ve probably experienced the pain of complex dev envs too much and have shaped my dev style around using the simplest possible environments).
What I’ve been doing lately is keeping a bash script that records the steps I take when setting up my environment. I guess the eventual hope for me is to do all my dev in a fresh debian vm, and be able to run dev_env_setup.sh from the repo I’m working on, then at the end of coding session push code changes and destroy the vm. This ensures the script is run often so is kept up to date. But… I was hoping to use something lighter than VMs. I have the scripts but they act mainly as a kind of note taking rather than a tool I actually use in practice.
My main motivation for taking this direction is I really do not want to run nodejs locally (I can’t even justify it rationally so don’t ask, it’s a visceral thing) so I decided to begin taking steps toward only developing in ‘temporary’ dev environments. Not there yet because it’s less work to do everything locally like I am now, but that workflow lends itself to the idea of nix style deterministic environments. This is also a very normal and simple workflow in qubes, but like my feeling on VMs it feels a little heavy and the overhead is hard to justify.
On Nix, I’m thinking more about making it easier for anyone to build stuff rather than to make my life easier. Like you I’m mostly over the pain, but it still wastes time. If I could just run a Nix config, or publish it for anyone wanting to build my stuff it could get more people involved and save them a lot of time. Like you I keep a note (but manually) whenever I build something new to make it easier to share our repeat the steps needed.
In the need for reproducible builds, I’m pretty sure this is the guy, although I only read his tweet dialog at the time:
https://blog.cmpxchg8b.com/2020/07/you-dont-need-reproducible-builds.html?m=1
And here’s a HN thread which I’ve not read either!
If you read I’d be interested to know if you agree or disagree on any important uses.
The article has some really good points.
To me reproducible builds is about setting an expectation for the degree of communication and quality of the software. A reproducible build means there must be a certain build process specified, it must be communicated in a useful way, the benefit vs cost must be considered.
That’s a bit of a soft answer, and those things can be achieved without reproducible builds. But an expectation of reproducible builds sort of ‘enforces’ a certain degree of this, whereas documentation alone or bash scripts alone can drift and decay with no clear indication of how much drift has actually happened.
I feel like the degrees of benefit are ignored (it’s not a binary yes benefit or no benefit). It’s better to be able to have ten separate individual builds which conform than a single one. Even if those ten could be colluding for malicious purposes or whatever, it seems like the degree of benefit is higher than an ad-hoc build.
I also feel like the case they present where ‘build it yourself means the external reproducible build becomes redundant’ is incorrect. Doing the build process yourself adds a data point of trust for the other people who have also done the build (and shared their results). A reproducible build encourages a group experience of the build process, whether doing the build or simply consuming the outputs as an end user. I feel there’s value in the social dynamic enabled by reproducible builds that can’t as easily be had with ad-hoc builds.
But, overall, I really appreciate the arguments put forward against reproducible builds. Thanks for the link.
You’re welcome @mav, and make useful valid points. A lot flows from the nuance, so good for me to look out for that in other contexts to. Thanks very much.
I just searched on the forum for mentions of Nix or NixOS. I’ve been using NixOS for almost two years and not thinking about switching to anything else. I’ve grown really fond of it for my local system, but also see potential for software packaging and cloud usage.
The only thing I realized a while back, is that with Nix you can’t really package it for other systems. This is something I have run into a few times before, where I was building the software locally and the binaries I produce will have references to libraries with ‘fixed’ paths on my system in the nix store (/nix/store/<package>/). This can be solved, but it’s something inherent to Nix/NixOS.
One nice thing though is that Nix can build Docker containers that contain just the binaries and libraries needed to run e.g. a binary. Therefore you essentially have reproducible containers. (Which is an often heard critique on Docker, that after a few years, the images are outdated and hard to use or update.)
What I personally love is how I have configurations for my shell and developer environment that will allow me to have the exact same experience and tools on both my desktop and Macbook. And for a personal Raspberry Pi I can also re-use some stuff. It’s just super nice to have my complete setup located in a single repository.
I’ve just discovered the safe network, still poking around to see if it’s the right fit for my app–so I have a lot of ignorance re: context here. Sorry if this is out out place.
Here’s why I think nix is worth it:
Support is being added ([RFC 0133] Git hashing and Git-hashing-based remote stores by Ericson2314 · Pull Request #133 · NixOS/rfcs · GitHub) which will let nix users reference packages by git hash instead of by the URL of github/gitlab/whatever. This excites me because it means that you can bring new determinism to code that might otherwise be problematic to reference in a distributed context, and you have a consistent way of referencing calls into it in a transparent, memoizable and partition-tolerant way (i.e. you don’t don’t have to talk to github as long as someone nearby has the bits).
You don’t have to be building software, you’re building consistent execution environments. You could be building proofs or simulation outputs or trained models.
Imagine a user that asserts somewhere:
fetch repo/ref
sha256:aaaaaaaaaaa, name it X
get data
sha256:bbbbbbbbb, name it in a file: Y
using the root repo of X do this:
nix run Y
hash stdout from the above command, you’ll get
sha256:ccccccc
So that’s a tuple of hashes representing a function call: ("aaaaaaaaaaa", "bbbbbbbbb", "ccccccc") and users can verify it by running the above steps (in a sandbox which becomes network-disabled as soon as nix is done fetching its inputs, which are hash-pinned in the flake.lock in the repo).
If someone you trust has already verified that the function defined in the tuple is pure and indeed has ccccccc as the output (given bbbbbbbbb) you can skip verification it and just fetch the output directly. Otherwise, you have everything you need to verify it so that your friends don’t have to.
One can imagine using this to create a mesh of memoized verifiable function calls which derive some useful result, and that’s what my app is about.
So that was a lot of words, sorry. My point here is that repeatable builds have utility between users of distributed systems, even if those systems are running on components that are not built in a repeatable way.
For me, that makes it worth embracing repeatable-builds life everywhere, because once you’ve paid the initial cost of figuring nix out (which, admittedly, is high) you might as well benefit from it everywhere that you can.
Besides, your developers who don’t like nix can always just ignore it continue configuring their system to match what they’re developing. You don’t have to declare the flake.nix authoritative until you’re convinced it’s actually useful.