A good example of 'use the right tool for the right job'. Nix is great for some use cases and awful for others. The problem is the Nix learning curve is so high that by the time you've grasped it enough to make a decision you feel you've invested too much time to back out now and pivot to something else so you try to shoehorn it to solve the original need.
Ie it is very easy for an AI to create a to-spec shell.nix (some Python packages, some Linux packages, some env vars, some path entries etc), or configuration.nix because of this paradigm.
I do this a lot to include envs with repos that fully support the package. It would probably be more reproducible with flakes (a flake.nix is like a shell.nix but with version pinning… or something, I’m still climbing that learning hill).
> Railway injects a deployment ID environment variable into all builds.
They could've done it in the next layer after installation. Also, you can split packages into different layers. There's even automation for it if you need batches to keep the number of layers down.
Care to elaborate what that means and what the alternative is?
Package maintainers often think in terms of constraints like I need a 1.0.0 <= pkg1 < 2.0.0 and a 2.5.0 <= pkg2 < 3.0.0. This tends to make total sense in the micro context of a single package but always falls apart IMO in the macro context. The problem is:
- constraints are not always right (say pkg1==1.9.0 actually breaks things)
- constraints of each dependency combined ends up giving very little degrees of freedom in constraint solving, so that you can’t in fact just take any pkg1 and use it
- even if you can use a given version, your package may have a hidden dependency on one if pkg1’s dependencies, that is only apparent once you start changing pkg1’s version
Constraint solving is really difficult and while it’s a cool idea, I think Nixpkgs takes the right approach in mostly avoiding it. If you want a given version of a package, you are forced to take the whole package set with you. So while you can’t say take a version of pkg1 from 2015 and use it with a version of pkg2 from 2025, you can just take the whole 2015 Nixpkgs and get pkg1 & pkg2 from 2015.
Thank you, I was looking for an explanation of exactly why I hate Nix so much. It takes a complicated use case, and tries to "solve" it by making your use-case invalid.
It's like the Soylent of software. "It's hard to cook, and I don't want to take time to eat. I'll just slurp down a bland milkshake. Now I don't have to deal with the complexities of food. I've solved the problem!"
It removes the “magic” constraint solving that seemingly never works and pushes it to the user to make it work
Note that the parent said "I think Nixpkgs takes the right approach in mostly avoiding it". As others have already said, Nix != Nixpkgs.
If you want to go down the "solving dependency version ranges" route, then Nix won't stop you. The usual approach is to use your normal language/ecosystem tooling (cabal, npm, cargo, maven, etc.) to create a "lock file"; then convert that into something Nix can import (if it's JSON that might just be a Nixlang function; if it's more complicated then there's probably a tool to convert it, like cabal2nix, npm2nix, cargo2nix, etc.). I personally prefer to run the latter within a Nix derivation, and use it via "import from derivation"; but others don't like importing from derivations, since it breaks the separation between evaluation and building. Either way, this is a very common way to use Nix.
(If you want to be even more hardcore, you could have Nix run the language tooling too; but that tends to require a bunch of workarounds, since language tooling tends to be wildly unreproducible! e.g. see http://www.chriswarbo.net/projects/nixos/nix_dependencies.ht... )
You could not depend on a patch version directly in source. You could force a patch version other ways, but each package would depend on a specific major/minor and the patch version was decided at build time. It was expected that differences in the patch version were binary compatible.
Minor version changes were typically were source compatible, but not necessarily binary compatible. You couldn’t just arbitrarily choose a new minor version for deployment (well, you could, but without expecting it to go well).
Major versions were reserved for source or logic breaking changes. Together the major and minor versions were considered the interface version.
There was none of this pinning to arbitrary versions or hashes (though, you could absolutely lock that in at build time).
Any concept of package (version) set was managed by metadata at a higher level. For something like your last example, we would “import” pkg2 from 2025, bringing in its dependency graph. The 2025 graph is known to work, so only packages that declare dependencies on any of those versions would be rebuilt. At the end of the operation you’d have a hybrid graph of 2015, 2025, and whatever new unique versions were created during the merge, and no individual package dependencies were ever touched.
The rules were also clear. There were no arbitrary expressions describing version ranges.
This is likely the source of their commit based versioning complaint/issue, i.e the commits in question are probably https://github.com/NixOS/nixpkgs versions if they aren't maintaining their own overlay of derivations.
This is in contrast to systems that allow all of the versions to move independently of each other.
i.e in the Nix world you don't just update one package, you move atomically to a new set of package versions. You can have full control over this by using your own derivations to customise the exact set of versions, in practice most folk using Nix aren't deep enough in it for that though.
Packages in nixpkgs follow the "managed distribution" model, where almost all package combinations can be expected to work together, remain reasonably stable (on the stable branch) for 6 months receiving security backports, then you do all your major upgrades when you jump to the next stable branch when it is released.
It also puts a function in the result, called `override`, which can be called to swap out any of those arguments.
"Default versions" breaking things that depend on them? What is that? It is like using docker's ":latest" tag and being surprised each time that a new server falls on its face because the "default" image is actually a different version from the previous "default" image.
I don't understand any of the explanations in this blog post. Seems like people who have zero clue about what a "version" of a software is.
"no way of splitting up the Nix dependencies into separate layers" - Why? Of course you can split /nix/store into as many layers as you need. Do they even know how to use containers and how to use Nix in the first place?
With the clear incompetence of these people, no wonder that their proposed solution smells like a decomposed fish.
Classic NIH syndrome. There is going to be no surprise to see them meet the exact same problems they didn't solve with Nix to infest their new "solution".
A nix wrapper or a deployment platform
I don't think any VC worth the time is going to sit around nitpicking how much Nix matters to their offering if they're making increasing amounts of money.
It sounds like it's a little bit too late, but I'm happy to provide some consulting on how you can get it to work idiomatically with Nix.
Product looks cool!
sure you are not going to get shared library conflicts, but i think this solution is extremely wasteful, and can make development painful too - look at nixpkgs' staging process.
Each person doesn't have to perform the build on their own. A build server will evaluate it and others will pull it from the cache.
The greater waste that nix eliminates is the waste of human time spent troubleshooting something that broke in production because of what should have been an innocent change, and the lost business value from the decreased production. When you trust your dependencies are what you asked for, it frees the mind of doubt and lets you focus on troubleshooting more efficiently towards a problem.
Aside, I spent over a decade on Debian derived distros. I never once had one of these distros complete an upgrade successfully between major versions, despite about 10 attempts spread over those years, though thankfully always on the first sacrificial server attempted. They always failed with interesting issues, sometimes before they really got started, sometimes borking the system and needing a fresh install. With NixOS, the upgrades are so reliable they can be done casually during the workday in production without bothering to check that they were successful. I think that wouldn't be possible if we wanted the false efficiency of substituting similar but different packages to save the build server from building the exact specification. Anything short of this doesn't get us away from the "works on my machine" problem.
Of course lots of software isn't ready for but reproduction which is why Nix has taken such a pragmatic approach. (I have written a lot about this).
It's all a series of tradeoffs. If your goal is reproducibility (as close as you can get), you will have a larger graph likely ..since you are accounting for more!
Sometimes we like to believe we can have our cake and eat it too rather than understand life's a series of tradeoffs.
When we think we are getting a silver bullet, we've likely just pushed that complexity somewhere else.
> When we think we are getting a silver bullet, we've likely just pushed that complexity somewhere else.
True but we kind of just stopped looking. and I feel much of the solution space hasn't been explored.
I totally understand the value proposition of Nix. However I think saying "bad time" is a bit hyperbolic. At most it's "You'll be losing a pretty significant guarantee compared to Nix". Still probably "packed to be more likely to work correctly" than 95% of software out there.
Go is the best choice at the moment for such tools. These tools start a process, do lots of IO and exit.
Very pragmatic choice.
Meanwhile Rust requires a pile of variable quality community driven crates to do basic things.
And Go's runtime is built-in by default. Unlike Java so there's nothing to "carefully set".
This is not done by default to reduce binary sizes.
Go can cross-compile from Linux to Windows, Darwin and FreeBSD without requiring any external tooling.
> The biggest problem with Nix is its commit-based package versioning. Only the latest major version of each package is available, with versions tied to specific commits in the nixpkgs repo.
While Nixpkgs is an amazing resource, Nix != Nixpkgs. Nixpkgs is highly unideal for cases where you want to be able to pull arbitrary versions of toolchains, but it is not the only way to go. For example, there is amazingly good Nix tooling for pulling an arbitrary version of Rust. Other Nix-based developer tools have shown how you can do this well.
> no way of splitting up the Nix dependencies into separate layers
That doesn't make any sense. You can literally just split them into separate layers in whatever arbitrary fashion you'd like. The built-in Nixpkgs docker tooling has some support for this even.
> We also changed the codebase from Rust to Go because of the Buildkit libraries.
This part is not related to Nix, but I find it interesting anyways. Obviously most people don't transition programming languages on a whim, it's generally something you do when you're already planning on building from scratch anyways. To me it almost sounds like different people worked on Railpacks vs Nixpacks.
(I've definitely seen what happens when people not familiar with Nix wind up having to deal with unfinished Nix solutions within an organization. It is not pretty, as most people are unwilling to try to figure out Nix. I don't generally use Nix at work out of fear of causing this situation.)
> While Nixpkgs is an amazing resource, Nix != Nixpkgs.
If Nixpkgs is the default and alternatives require additional research and effort then for most users it _is_ Nix.
> That doesn't make any sense. You can literally just split them into separate layers in whatever arbitrary fashion you'd like. The built-in Nixpkgs docker tooling has some support for this even.
Is this obvious, simple, and default behaviour?
We are talking about a company full of professionals. If they need something obvious, simple, and default to manage their build - the core business function that turns their text into deployable artifacts - maybe there is a skill culture issue.
The industry is full of ineptitude though.
While I disagree with the person you're replying to, I find your reply dismissive.
I don't know the behind-the-scnene reasons for this, but I can very very easily apply a very similar situation to this from my experience.
Nix is a full blown functional programming language along with a very rich (and poorly documented, niche, only second to C++ template in error comprehensibility[1]) ecosystem in itself. It's not like "docker" or "kubernetes" where you're mostly dealing with "data" files like yaml, json or Dockerfile. You're dealing with a complex programming project.
With that in mind:
- You have a core team with 1 or 2 people with Nix passion/expertise.
- Those people do most of the heavy lifting in implementation.
- They onboarding the team on to Nix
- They evangelize Nix through the org/company
- They mod and answer all the "#nix-discussions" channel questions
Initially the system is fairly successful and everything is good. over the next 5-6 years it would accumulate a lot of feature asks. The original "Nix person" has long left. Most of the original people have moved either to other projects or not particularly that passionate about Nix. In fact, the "best" developer you have who has inherited the whole Nix thing has only really had to deal with all the shit parts of Nix and the system. They are they ones fixing issues, dealing with bugs, etc. All while maintaining 3 stacks, a Nix stack, a Go stack, and a Rust stack.
Eventually that person/team that's annoyed by maintaining the Nix project wins. They want to own that code. They don't want to use Nix any more. They know what's needed, they want to implement it as part of their main Go stack that they are actively working on. They can optimize things for their special case without having to worry about "implementing it the Nix way" or "doing it upstream".
They promise you (the management who is open to the idea, but trying to understand the ROI) feature parity + top 5 feature asks for the initial release. You trust the team enough to let them do what they think is best.
[1]: LLMs are really good at suggesting a solution given an error message. Nix errors bring them to their knees. It's always "Hmmm.... it appears that there is an error in your configuration... have you tried a `git revert`?"
Just because people decide stuff for money doesn't mean I can't call them bad. Not everyone is equally skilled.
And your parable is exactly the issue. The unskilled and loud and whiny do often win, and it's a shame. I see it all the time.
(Also you're way overstating Nix as a "full blown FP language." It isn't hard to learn. I learned it just be existing on a project with it. Within 6mo, now I'm apparently a "Nix expert" and people now point at me as one of the people who "knows it" and "you can't expect everyone to know it like you do." idk maybe I'm some genius but I think it's more that I just don't have a bad personality.)
> A perfectly capably (but perhaps a bit esoteric) technology is picked by a smart passionate person for a project.
> The novel technology is in 1 isolated module that's mostly feature complete for the first 1-3 years.
> People in the team/company deal with that "thing" as a blackbox more and more
> 5-10 years later, mostly new team maintaining the project. They hate the weird choice of tech. "Why is only this one component different???"
> People understand the contract with the "black box" very well, but HATE the "black box". People think "We can implement the black box contract very easily"
And I am dismissing the types you describe specifically. I dismiss them (privately amongst the likeminded) at work all the time too. I just put them on a list in my head when they start spouting these sorts of bad values.
This feels rather dismissive. They wrote a bespoke solution, not a weekend toy. Surely you'd agree that they have more than just surface-level knowledge of Nix, to be able to distinguish between Nix and Nixpkgs? They're already doing non-trivial things by merging multiple commits of Nixpkgs in order to get different versions of different tools!
> Is this obvious, simple, and default behaviour?
Well, Nix doesn't do much of anything "by default", it's a pretty generic tool. But insofar as it matters, Yes, pretty much. `dockerTools.buildLayeredImage` will in fact automatically build a layered image, and it is the most "obvious" way (IMO) to build a docker image. There is also `dockerTools.buildImage` but there's no particular reason to use it unless you specifically want a flattened image. (The documentation available is clear enough about this. In fact, in practice, much of the time you'd probably actually want `dockerTools.streamLayeredImage` instead, which is also documented well enough, but that's beyond the point here.)
But that's not my point. As far as I know, Nixpacks don't even use this functionality, I'm pretty sure they wrote their own OCI image building tools. And in that sense, it is not obvious why they can't split the Nix store and the article doesn't explain it.
My point wasn't to be dismissive about the difficulties of Nix, it's that the blog post doesn't really do a good job of explaining things. It makes it sound like these are normal problems in Nix, but they are not; even the official Nixpkgs documentation often points to third party solutions for when you're working outside of Nixpkgs, since most of the Nixpkgs tools is geared for Nixpkgs and NixOS usage. As an example, take a look at this section of the Rust documentation in Nixpkgs:
https://github.com/NixOS/nixpkgs/blob/master/doc/languages-f...
So even if you're relatively new to Nix, as long as you are reading the documentation you will indeed definitely be aware of the fact that there is more to the Nix ecosystem than just Nixpkgs. It may not be surface-level Nix knowledge, but it's certainly close.
If you identify these things as an issue, any competent engineer should find a variety of solutions with search and/or LLM assistance within an hour, since they’re not super obscure requirements.
I’m not saying Railway didn’t do this and realize that these common solutions weren’t viable for them, but it’s odd to not mention anything they tried to get around it.
But as kfajdsl points out: that's not what TFA is. This is a company building a product on top of Nix. Package management is their expertise. Anyone using Nix in that capacity understands the distinction between nix and nixpkgs. Which they certainly do--GP only remarked it was odd they didn't explain it, not that they didn't know.
The Nix language is something I could criticize for hours without getting bored, but it is what it is. It's old and they did the best they could and it's probably not worth changing. The Nix build system feels awfully primitive to me, often rebuilding stuff that doesn't need to be rebuilt for no good reason. (For example, my NixOS installer ISO has a ton of the build depend on the cmdline I pass to the kernel [just console=ttyS2,1500000n8], and so changing the speed of my serial port requires about 3 minutes of build time. It's goofy and makes me laugh, I'm not going to stop using Nix because of it... but it's also something that I wouldn't let happen in MY build.)
Nix for Docker images is, in my opinion, what it's the worst at. A long time ago, I was writing some software in Go and needed to add the pg_dump binary from Postgres to my container image. The infrastructure team suggested using Nix, which I did, but our images blew up from 50MB of our compressed go binary to 1.5GB of God Knows What. pg_dump is 464K. I ended up doing things my way, with Bazel and rules_debian to install apt packages, and the result (on top of distroless) was much cleaner and more compact. My opinion with some actual Nix experience is that a Nix system always ends up being 1.4GB. My installer ISO is 1.4GB. My freshly installed machine is 1.4GB. That's just how it is, for whatever reason.
Finally, the whole "I would like to build a large C++ project" situation is a well worn path. s/C++/Rust doesn't change anything material. There are build systems that exist to make the library situation more tolerable. They are all as complicated as Nix, but some work much better for this use case. Nix is trying to be a build system for building other people's software, supporting nixpkgs, and lands on the very generic side of things. Build systems that are designed for building your software tend to do better at that job. Personally, I'm happy with Bazel and probably wouldn't use anything else (except "go build" for go-only projects), but there are many, many, many other options. 99% of the time, you should use that instead of Nix (and write a flake so people can install the latest version of Your Thing with home-manager; or maybe I'm just the only person that uses their own software day to day and you don't actually need to do that...)
That's strange, I never had problems building really tiny docker (release) images with nix, in fact it felt easier than doing it with alpine. You just get exactly what you specify, no more.
(OTOH, when developing in nix, I always end up with a huge /nix/store and have no idea how to clean it without garbage collecting everything and having to wait all over)
FYI you can avoid things getting garbage-collected by doing `nix-store --add-root`; that makes an "(indirect) garbage collector root"[0]. Especially useful if you're using import-from-derivation, since that imported derivation won't appear in the dependencies of your final build output (which, to be clear, is a good thing; since it lets us calculate a derivation, e.g. by solving dependency constraints or whatever, without affecting the eventual hash if that calculation happens to match a previous one!)
[0] https://nix.dev/manual/nix/2.18/package-management/garbage-c...
> Smaller Builds > Better caching
what were the benefits that overcame this, and what about those now?
I am naive about Nix, but...
...isn't that like...the whole selling point of Nix? That it's specific about what you're getting, instead of allowing ticking time bombs like python:latest or npm-style glibc:^4.4.4
Odd to attach yourself to Nix then blog against its USP.
nix2container [1] is actually able to do that: you can explicitly build layers containing a subset of the dependencies required by your image. An example is provided in this section: https://github.com/nlewo/nix2container?tab=readme-ov-file#is...
For instance, if your images use bash, you can explicitly create a layer containing the bash closure. This layer can then be used across all your images and is only rebuild and repushed if this bash closure is modified.
> > pull in dependencies often results in massive image sizes with a single /nix/store layer
This is the case for the basic nixpkgs.dockerTools.buildImage function but this is not true with nix2container, nor with nixpkgs.dockerTools.streamLayeredImage. Instead of writing the layers in the Nix store, these tools build a script to actually push the image by using existing store paths (which are Nix runtime dependencies of this script). Regarding the nix2container implementation, it builds a JSON file describing the Nix store paths for all layers and uses Skopeo to push the image (to a Docker deamon, a registry, podman, ...), by consuming this JSON file.
(disclaimer: i'm the nix2container author)
Completely different approach to dependencies, though. For now.
The OS should be immutable. Apps and services and drivers/extensions should be self contained. Things should not be installed “on” the OS. This entire concept is a trillion dollar mistake.
lloydatkinson•8h ago
> This approach isn’t clear or maintainable, especially for contributors unfamiliar with Nix’s version management.
> For languages like Node and Python, we ended up only supporting their latest major version.
What is not maintainable about this? That they need to make a list of available versions? So, can this not be automated?
Furthermore, why is Railway defining how a user uses Nix?
Surely one of the points of Nix is that you can take a bare machine and have it configured with exactly what versions of packages you want? Why would Railway need to get in the way of the user and limit their versions anyway?
Or did I misunderstand and they don’t even expose Nix to the user? If so, the original question still stands: can’t they automate that list of package versions?
codethief•8h ago
That's at least my understanding, yes.
notnmeyer•8h ago
elbear•7h ago
Honestly, the reasons given don't feel very solid. Maybe the person who introduced Nix left and the ones remaining didn't like it very much (the language itself is not very nice, the docs weren't great either in the past).
Still, I'm not familiar enough with the stack they chose, but does it provide a level of determinism close to Nix? If not, it might come to bite them or make their life harder later on.
e3bc54b2•7h ago
And yes, their reasoning implies NIH and just unfamiliarity combined with unwillingness to really understand Nix.
[0]: https://discourse.nixos.org/t/how-long-is-binary-cache-kept-...
[1]: https://hal.science/hal-04913007
[2]: https://luj.fr/blog/is-nixos-truly-reproducible.html
[3]: https://discourse.nixos.org/t/nixos-foundations-financial-su...
[4]: https://discourse.nixos.org/t/the-nixos-foundations-call-to-...
mplanchard•4h ago
akho•5h ago
koolala•5h ago
akho•4h ago
cormacrelf•7h ago
> The way Nixpacks uses Nix to pull in dependencies often results in massive image sizes with a single /nix/store layer ... all Nix and related packages and libraries needed for both the build and runtime are here.
This statement is kinda like “I’m giving up on automobiles because I can’t make them go forward”. This is one of the things Nix can do most reliably. It automates the detection of which runtime dependencies are actually referenced in the resulting binary, using string matching on /nix/store hashes. If they couldn’t make it do that, they’re doing something pretty weird or gravely wrong. I wouldn’t even know where to start to try to stop Nix from solving this automatically!
I wouldn’t read too much into their experience with it. The stuff about versioning is a very normal problem everyone has, would have been more interesting if they attempted to solve it.
mplanchard•4h ago
Image size isn’t something we’ve focused a lot on, so I haven’t spent a ton of time on it, but searching for “nix docker image size” shows it to be a pretty commonly encountered thing.