Developing on NixOS

I've now been using NixOS for over 6 months, and it's finally starting to make sense and become second-nature to me. I didn't have any prior experience using Nix, so it's been quite a learning curve to base my whole OS on it!

Here's a rough guide to how I've been getting along, in chronological order from "struggling to make sense of what these blog posts and Wiki pages say", up to "even my one-liner utility scripts are integrated into Nix now".


This was the first clear benefit. Nix packages provide various hooks into the software they contain, which lets us alter their configuration from one place (/etc/nixos/configuration.nix for system-wide things like Grub, Linux, X, etc. and ~/.nixpkgs/config.nix for per-user settings).

Also, since this is all integrated into Nix, new packages will be installed automatically when our configuration happens to use them (eg. designating a display manager for X), plus all changes are atomic and we can roll back to previous configurations whenever we want (past system configurations are also available via Grub, in case your new config doesn't boot).


After getting to grips with configuration.nix, I carried on my normal approach of installing packages as and when I needed them. I basically replaced my usual habit of running apt-get install foo with running nix-env -i foo instead.

Internally, Nix gives every user their own "profile", which is just a Nix package. Running nix-env -i foo will actually create a new version of my profile, with foo added to its dependencies. For convenience, this new profile will be installed immediately and the symlinks in ~/.nix-profile/ will be updated accordingly, so I will be able to run foo straight after nix-env has finished.

There are few benefits to nix-env over apt-get:

  1. Packages are installed per-user, rather than system-wide
  2. Multiple versions can be installed side-by-side
  3. I can roll back to previous profiles whenever I like

Despite these advantages, there are actually much better ways of using NixOS.


I spend a lot of time developing software, so I usually have a load of little helper scripts living in /usr/local (these are usually symlinks to a git repository somewhere in ~).

Doing this on NixOS would undermine the benefits of Nix, so I looked for a way to add them gracefully.

The Nix package repository is maintained on Github so with a bit of effort we can get a copy and see how the packages work.

The -f option of the nix-env command allows us to specify our own copy of nixpkgs

This lets us add our own packages to a local clone of nixpkgs, without having to push these packages upstream. Since nixpkgs is all in git, we can even maintain our changes in separate branches if we want.

The key to using custom software in a Nix package is the package's src. This specifies where the code will come from, and among other things it allows URLs, VCS repositories and local file paths. If your code is maintained in a VCS, then you might as well specify that. If it's stuff you'd rather not put online, then use a file path (remember that you can use git and co. offline too!)

The problem with using a separate nixpkgs for custom code is that there is a separation between the software itself and its Nix dependency information. Of course, the ability to maintain them separately is vital, but since we're maintaining both, it would be nice to keep them in lock-step.

Also, whilst using a different tree for a few nix-env commands is useful for resolving dependencies between multiple custom packages, Nix will not have access to these custom definitions at any other time; unless we use our custom nixpkgs clone as the system's default repo, which imposes a maintenance burden.


To keep package definitions in the same source repository as our software, we just need to put them in a default.nix file. If we do this, we gain a remarkable new ability: by running nix-shell from the directory containing default.nix, we're dropped into a shell with all of the specified dependencies available. When we exit the shell, they're free to be garbage-collected.

In other words: we don't need to install the package in order to use it!

I've been making extensive use of default.nix files (and, if there are a bunch of nix-shell-specific things that don't belong in the regular package, the file shell.nix can also be used). These ensure a clean, consistent, reproducible development/test/deployment environment.

In particular, we can use nix-shell --pure to ensure that only the specified dependencies are available. For example, if your user profile has bash available, but it's not listed as a dependency in default.nix, then using nix-shell without --pure will cause bash to be available inside the resulting shell.

This may be convenient for one-off tasks (eg. if you want to launch an editor from inside the shell to edit the code in a particular way, but don't want to add it as a dependency), but it's not a good idea for frequent or important work, since that software may become an implicit dependency without you realising it. For example, imagine if you only ever run your test suites with nix-shell, not nix-shell --pure; your tests might fail without some unspecified dependency, but you'd never know!

Instead, we should strive to always use nix-shell --pure; that way, if we try to use something which isn't listed as a dependency, we will get an error (eg. to install bash, run nix-env -i bash). We can then add that dependency to default.nix to ensure that it will always be available whenever our software is used.

Continuous Integration

Since writing this page, I've built up a new habit which is analogous to continuous integration in the world of online services. Namely, I tie my Nix packages to their git repositories, such that the package source is always the latest commit, and a simple git push can make any change instantly available to nix-env or nix-shell.


That's about as far as I've got with Nix/NixOS so far. If I develop some better habits, I may update this page with more info!