Technical groups
Dropdown arrow
Open source
Consulting services
Technical groups
Dropdown arrow
Open source
Consulting services

Optimizing NixOS Search

13 December 2022 — by William Wang

With their introduction in Nix 2.4, flakes are quickly becoming an integral part of the Nix ecosystem. For anyone unfamiliar, flakes exist to provide a standard format to package Nix-based projects.

To allow for better user experience and composability for existing flakes, discoverability of flakes is a necessary feature.

The site is used to search for packages, options, and flakes. It also provides metadata information such as descriptions, maintainers, and compatible platforms.

Currently, imports flake metadata through custom nix code that wraps flake evaluations. However, wrapping employs too much memory when used with large repositories such as nixpkgs. This causes a segmentation fault. As currently implemented, cannot show all flakes.

My internship involved demonstrating a solution to this issue by updating the JSON schema for nix flake show to natively export the necessary metadata for packages.

Metadata For Export

nix flake show reads in the file flake.nix which provides an interface to artifacts like packages. Packages contain metadata fields (also called meta-attributes) detailing information such as a description of the package (meta.description) and a list of supported Nix platforms (meta.platforms). To showcase a working solution, I decided to implement support for three additional fields: meta.homepage, meta.licenses, and meta.maintainers.

These fields are more or less how nixpkgs defines them in their flake files. In theory, these changes to nix flake show should already be compatible.

Let’s explore these three additional fields:

homepage is the package’s homepage. It takes the form of a simple string:

meta.homepage = "";

licenses appears as an attribute set:

meta.licenses =
    spdxId = "MIT";
    fullName = "MIT License";

maintainers includes a list of package maintainers and is defined as a list of attribute sets which include basic identifying information:

meta.maintainers =
            email = "[email protected]";
            matrix = "";
            name = "first last";
            github = "username";
            githubid = 0000000;
            keys = {
                fingerprint = "8w5w jc2b d0h8 q3l9 0tc4 m55v i87r 23zj 0fbj r30x"

For example, the GNU Hello package declares its meta fields as follows:

meta = with lib; {
  description = "A program that produces a familiar, friendly greeting";
  longDescription = ''
    GNU Hello is a program that prints "Hello, world!" when you run it.
    It is fully customizable.
  homepage = "";
  license = licenses.gpl3Plus;
  maintainers = [ maintainers.eelco ];
  platforms = platforms.all;

Flake Show Command

The nix flake show command is implemented by CmdFlakeShow’s run() function. The run() function consists of a definition and call to the local function visit(). For recognized nested attributes, visit() will be recursively run until all recognized attributes have been processed. When applicable, another nested function showDerivation() will be run, printing a derivation’s attributes to standard output.

When the json flag is passed into the program, output information and metadata will also be stored in a JSON object. Upon recursing, showDerivation will check for a meta symbol within the flake.nix file. If found, the program will continue to check for recognized symbols defined as elements under meta. To collect our metadata, we run maybeGetAttr which returns std::shared_ptr<AttrCursor>. As AttrCursor serves as a wrapper for an Attr, we can use one of its member functions to access each metadata value based on its type.

For description and homepage, this is straightforward. AttrCursor implements a getString() function that returns a std::string value for a given attribute.

license consists of an attribute set. Therefore, we can use the getAttrs() function to obtain a std::vector<Symbol>. We can then do a lookup of the symbols in our EvalState for our std::string values.

So far, the previous two fields are easy to implement. However, maintainers requires a bit more work. The main trouble comes from the fact that maintainers is defined as a list of attribute sets. There exists no getter function for this attribute value type. This requires us to implement our own getListOfAttrs() function for AttrCursor.

Implementing getListOfAttrs

For performance reasons, Nix internally caches the evaluation results on-disk.

This cache uses a SQlite database with a custom schema designed for efficiently caching lazy recursive data-structures.

However, this format does not yet handle lists properly, and only has hard-coded support for lists of strings, which is not sufficient for our purposes, as some of the metadata fields are represented as lists of compound structures.

Adding such support is easy in principle, but the actual implementation would have taken more time than what was available for my internship. My solution instead was to fall back on an uncached evaluation for evaluating these attributes. However, this means that we will not benefit from the cache, making it a temporary solution only.

Nevertheless, with this prototype solution, nix flake show now shows all three metadata fields as JSON. This can now be used with nixos-search to deprecate our wrapped nix evaluations.

So what does the end result look like? Let us take a look at what nix flake show --json outputs when given a flake.nix file using the GNU Hello package:

"homepage": "",
"license": {
  "deprecated": "false",
  "free": "true",
  "fullName": "GNU General Public License v3.0 or later",
  "redistributable": "true",
  "shortName": "gpl3Plus",
  "spdxId": "GPL-3.0-or-later",
  "url": ""
"maintainers": {
  "edolstra": {
    "email": "[email protected]",
    "github": "edolstra",
    "githubId": "1148549",
    "name": "Eelco Dolstra"

Future Work To Be Done

  • As mentioned above, getListOfAttrs() should utilize caching in production. This will require revamping how lists are internally represented. Ideally, there should be a generalized getList() function that behaves similarly to getAttrs(). One proposed solution is to implement lists exactly like attribute sets but with a different tag and increasing integer key values.
  • There are many more metadata fields that can be queried. Implementing support for these in nix should be relatively straightforward.
  • nix flake check should warn users about missing or improperly formatted metadata fields.
  • Finally, we should implement our solution at and benchmark it to verify that the import of nixpkgs is fast enough.

Interning at Tweag

As my first internship experience, Tweag has been a fantastic work environment. Tweagers share a mutual love for discussion and group learning, which fosters the exploration and application of new ideas.

During my time here, I have had the chance to put common software engineering techniques into practice and immerse myself in all things Nix. My mentor Rok provided extensive advice throughout this process and taught me the value of avoiding premature optimizations.

My chapter at the Tweag Paris office gave me an inside glimpse into the brilliant minds behind all those GitHub commits. I would like to thank Tweag for this opportunity and my coworkers for their thought-provoking insights.

For a chance to experience working with a company where research and engineering intersect, I would highly recommend checking out Tweag.

About the author

William Wang

If you enjoyed this article, you might be interested in joining the Tweag team.

This article is licensed under a Creative Commons Attribution 4.0 International license.


AboutOpen SourceCareersContact Us

Connect with us

© 2024 Modus Create, LLC

Privacy PolicySitemap