Nix: Haskell Concepts for Package ManagementMay 21, 2018
Back in my BayHac article, I discussed some of my adventures with Nix and HNix. I didn’t get a lot done. But I was still curious to learn more about these systems. I “used” Nix a little bit at a previous job. And by “used” I mean I learned enough of the basic commands to write code and get on with my life. But I never developed a full understanding of “why Nix” or “what’s good about Nix”. So I’m going to spend a couple weeks doing a high level overview of this program and why it’s so cool.
As an introduction, Nix is a purely functional package manager. It aims to be a language-agnostic system to achieve deterministic builds. We’ll get into what it means to be a “purely functional” package manager down below. But a lot of the properties that make Nix what it is are also present in Haskell. So while you could use Nix for any language, most of the development effort so far has come from Haskellers. Meanwhile, NixOS is a linux distribution that seeks to apply the main principles of Nix at the operating system level.
This first article will discuss the basics of Nix, its advantages, and disadvantages. Next week, we’ll take a look at the HNix project, which seeks to implement Nix in Haskell. It’s important to understand though that Nix is definitely not the easiest package manager to use for Haskell. For now, I wold still recommend starting out with Stack. You can read the docs or check out our free Stack mini-course to learn more! And if you’ve never used Haskell before, download our Beginners Checklist to get started!
Now to motivate the use of Nix, let’s consider some of the broader issues are with package management.
At the most basic level, a package manager should enable you to get a program up and running in a small number (~3) of commands. And most accomplish this task, but there are always complications. We’ll look at two main issues. One is versioning. This includes both versioning your own projects and versioning dependencies. The other problem relates to the portability of your application.
The versioning problem plagued Haskell developers when Cabal was still young. Cabal would, by default, install dependencies system wide. But suppose you had many projects on your machine. These might depend on different versions of the same library. And this could lead to conflicts in your system that might render multiple projects unusable.
The addition of Cabal sandboxes and the Stack program mitigated this problem. Both these systems install dependencies in project specific locations. But there was still a problem where it could be difficult to roll back to a previous version of your project. The commands to uninstall and downgrade the packages weren’t intuitive. They could easily break things if you weren’t careful.
Meanwhile, unseen dependencies threaten our portability. This is somewhat more common in building C or C++ programs than Haskell programs. C libraries are often still installed system wide. One of the consequences is that you might have a library from another project on your system. Then a new project also depends on it, but you forget to list that dependency. It works fine for you on your local machine. But then when you push your code somewhere else, that dependency isn’t found. This can be quite a hassle.
The Nix Functional Approach
Nix (the package manager) seeks to avoid these problems by using a functional approach to package management. It treats every package as a value constructed by a function. The key input to the function of any package is its dependency graph. That is, a package is the final output, and the other (versioned) libraries are the input. Each version of a package you build has a unique identifier. This identifier is a cryptographic hash of the dependency graph. So if any of the dependencies to your program change, you’ll rebuild and create a totally new version of your package. This means adding dependencies, removing them, or changing versions.
Nix stores all its packages in the /nix/store directory. So you might build one version of your project that ends up in this directory:
And then you might change the dependencies and end up with another directory.
What are the consequences of this?
Notice it’s very easy to version our project! If we decide to rollback to a previous set of dependencies, that version will still be living on our machine! We’ll update the dependency set. It then calculates the hash of the dependency graph, and this will match an old configuration. So we’ll be all set! This goes for any of our dependencies as well.
There are in fact specific commands related to rollbacks. This means you can upgrade packages without being afraid of any difficulties.
Nix also solves the second problem we mentioned above. First, we explicitly declare all the dependencies as inputs. And second, we only use dependencies we get from the Nix store, rather than any system wide location. This means our derivations are complete. Thus someone else should be able to take the definition and build it themselves.
NixOS seeks to take many of the lessons from the Nix package manager and apply them at the OS level. Many of the problems that plague package management also plague OS management. For instance, upgrading packages with sudo apt-get install can be a risky operation. It can be difficult to rollback, and almost impossible to test what is going to happen before you upgrade. NixOS fixes these. It allows you to have versioned, reproducible system configurations. And you can roll back to a configuration with ease. It also gives you atomic transactions on system modifications. This way, even if something goes wrong, you’ll be completely reverted to your old system state.
Weaknesses with Nix
One potential weakness with Nix is that it defaults to building from source. This means you’ll often have long build times, even for small changes in your code or dependencies. If you’re in luck, you can use the Nix cache for your specific libraries. It stores pre-built binaries you can use. But from my experience using Nix, the length of build times was one of the biggest things holding it back. In particular it was very difficult to incorporate Nix into a CI system, as it was prone to cause timeouts.
So hopefully this gives you some idea of what Nix is about. Next week, we’ll look into HNix. This open source project is seeking to re-implement Nix in Haskell. We’ll see why in our exploration of the project. In the meantime, check out some of our resources on Getting Started with Haskell so you can learn how to get going! And if you want a little bit of experience with package management in Haskell, make sure to try out Stack! Check out our free Stack mini-course to learn how!