The plan is to replace it with dream2nix. The dream2nix python support is still WIP and needs your help. Please consider to support dream2nix with contributions or funds in order to speed up the process. Feel free to contact me.
Mach-nix makes it easy to create and share reproducible python environments or packages. Existing tools for python package management often suffer from reproducibility and complexity issues, requiring a multitude of tools and additional virtualization layers to work sufficiently. Mach-nix aims to solve these problems by providing a simple way to use nix, a revolutionary build system which is known to achieve great reproducibility and portability besides many other advantages.
- Users without nix experience, who want to maintain python environments for their projects which are reliable and easy to reproduce.
- Users already working with nix who want to reduce the effort needed to create nix expressions for their python projects.
- Use one tool to install packages from 3 different universes: pypi, conda, nixpkgs.
- Hardware optimizations, like for example SSE/AVX/FMA for tensorflow, are available without the need to manually mess with their build system. (see nixpkgs provider)
- Cross platform support (tested only aarch64)
- Easily include private packages or packages from other sources.
- Build time parameters and dependencies of complex python packages can be tweaked without needing to setup any build environment. It requires some knowledge about nix, though. For examples, see override system.
- mach-nix - Create highly reproducible python environments
- Table of Contents
nix-env -if https://github.com/DavHau/mach-nix/tarball/3.5.0 -A mach-nixor, if you prefer nix-shell:
-
nix shell github:DavHau/mach-nix
-
otherwise:
nix-shell -p '(callPackage (fetchTarball https://github.com/DavHau/mach-nix/tarball/3.5.0) {}).mach-nix'
mach-nix env ./env -r requirements.txtThis will generate the python environment into ./env. To activate it, execute:
nix-shell ./envThe ./env directory contains a portable and reproducible definition of your python environment. To reuse this environment on another system, just copy the ./env directory
and use nix-shell to activate it.
mach-nix gen -r requirements.txt...to print out the nix expression which defines a python derivation (optionally use -o to define an output file)
-
nix (build|shell) mach-nix#gen.(python|docker).package1.package2...
You can call mach-nix directly from a nix expression
let
mach-nix = import (builtins.fetchGit {
url = "https://github.com/DavHau/mach-nix";
ref = "refs/tags/3.5.0";
}) {};
in
mach-nix.mkPython {
requirements = ''
pillow
numpy
requests
'';
}find more examples under ./examples.md
Mach-nix can be fine tuned with additional arguments. Examples can be found in ./examples.md.
Functions for building python environments:
- mkPython - builds a python environment for a given
requirements.txt. - mkPythonShell - builds a python environment suitable for nix-shell.
- mkDockerImage - builds a layered docker image containing a python environment.
- mkNixpkgs - returns nixpkgs which conform to the given requirements.
- mkOverlay - returns an overlay function to make nixpkgs conform to the given requirements.
- mkPythonOverrides - produces pythonOverrides to make python conform to the given requirements.
Functions for building python packages or applications:
- buildPythonPackage - build a single python package from a source code while automatically detecting requirements.
- buildPythonApplication - same as buildPythonPackage, but package will not be importable by other python packages.
buildPythonPackage and buildPythonApplication accept the same arguments as their equally named partners in nixpkgs, plus the arguments of mkPython. If name/version/requirements arguments are omitted, mach-nix attempts to detect them automatically. See ./examples.md.
Note that some dependency declaration formats are missing. For a roadmap, please refer to issue #132.
mkPython and all other mk... functions take exactly the following arguments:
- requirements (string): Text content of a typical
requirements.txt.
- providers (set): define provider preferences (see examples below)
- packagesExtra (list) Add extra packages. Can contain tarball-URLs or paths of python source code, packages built via
mach-nix.buildPythonPackage, or R Packages. - _ (set): use underscore argument to easily modify arbitrary attributes of packages. For example to add build inputs use
_.{package}.buildInputs.add = [...]. Or to overwrite patches use_.{package}.patches = [...]. - overridesPre (list): (advanced) list of pythonOverrides to apply before the mach-nix overrides. Use this to include additional packages which can then be selected inside the
requirements - overridesPost (list): (advanced) list of pythonOverrides to apply after the mach-nix overrides. Use this to fixup packages.
- tests (bool): Whether to enable tests (default: false)
- _providerDefaults (set): builtin provider defaults. Disable them by passing {}
Providers allow you to configure the origin for your packages on a granular basis.
The following providers are available:
- conda: Provides packages from anaconda.org. Those packages can contain binaries. Some benefits of anaconda are:
- Different build variants for packages
- Provides all system dependencies for packages
- wheel: Provides all linux compatible wheel releases from pypi. If wheels contain binaries, Mach-nix patches them via patchelf to ensure reproducibility. Wheels are very quick to install and work quite reliable.
- sdist: Provides all setuptools compatible packages from pypi. It still uses nix for building, which allows it to be tweaked in a flexible way. But in some cases problems can occur, if there is not sufficient information available to determine required system depedencies.
- nixpkgs: Provides packages directly from nixpkgs without modifying their sources. Has only a few versions available, Use this provider if you need to tweak individual build parameters, like
SSE/AVX/FMAfor tensorflow for example.
Mach-nix builds environments by mixing packages from all 3 providers. You decide which providers should be preferred for which packages, or which providers shouldn't be used at all.
The default preferred order of providers is conda, wheel, sdist, nixpkgs.
Providers can be disabled/enabled/preferred like in the following examples:
-
A provider specifier like
"wheel,sdist,nixpkgs"means, that the resolver will first try to satisfy the requirements with candidates from the wheel provider. If a resolution is impossible or a package doesn't provide a wheel release, it falls back to sdist/nixpkgs for a minimal number of packages. In general it will choose as many packages from wheel as possible, then sdist, then nixpkgs. -
"nixpkgs,sdist"means, thatnixpkgscandidates are preferred, but mach-nix falls back to sdist.wheelis not listed and therefore wheels are disabled.
A full provider config passed to mach-nix looks like this:
{
# The default for all packages which are not specified explicitly
_default = "nixpkgs,wheel,sdist";
# Explicit settings per package
numpy = "wheel,sdist";
tensorflow = "wheel";
}Mach-nix will always satisfy the requirements.txt fully with the configured providers or fail with a ResolutionImpossible error.
If a mach-nix build fails, most of the time it can be resolved by just switching the provider of a package, which is simple and doesn't require writing a lot of nix code. For some more complex scenarios, checkout the ./examples.md.
Usually people rely on multiple layers of different package management tools for building their software environments. These tools are often not well integrated with each other and don't offer strong reproducibility. Example: You are on debian/ubuntu and use APT (layer 1) to install python. Then you use venv (layer 2) to overcome some of your layer 1 limitations (not being able to have multiple versions of the same package installed) and afterwards you are using pip (layer 3) to install python packages. You notice that even after pinning all your requirements, your environment behaves differently on your server or your colleagues machine because their underlying system differs from yours. You start using docker (layer 4) to overcome this problem which adds extra complexity to the whole process and gives you some nasty limitations during development. You need to configure your IDE's docker integration and so on. Despite all the effort you put in, still the problem is not fully solved and from time to time your build pipeline just breaks and you need to fix it manually.
In contrast to that, the nix package manager provides a from ground up different approach to build software systems. Due to its purely functional approach, nix doesn't require additional layers to make your software reliable. Software environments built with nix are known to be reproducible and portable, which makes many processes during development and deployment easier. Mach-nix leverages that potential by abstracting away the complexity involved in building python environments with nix. Under the hood it just generates and evaluates nix expressions for you.
The general mechanism can be broken down into Dependency resolution and Generating a nix expression:
Mach-nix contains a dependency graph of nearly all python packages available on pypi.org. This allows mach-nix to resolve dependencies offline within seconds.
The dependency graph data can be found here: https://github.com/DavHau/pypi-deps-db The dependency graph is updated on a daily basis by this set of tools: https://github.com/DavHau/pypi-crawlers
Despite this graph being updated constantly, mach-nix always pins one specific version of the graph to ensure reproducibility.
As core for the resolving resolvelib is used: https://github.com/sarugaku/resolvelib
Mach-nix supports multiple providers to retrieve python packages from. The user can specify which providers should be preferred. Packages from different providers can be mixed.
With pypi-deps-db, we have built a dependency graph, where each package is defined by name and version, for example pillow-9.1.0.
To download a package, we need it's URL and hash. This is where nix-pypi-fetcher-2 comes in.
nix-pypi-fetcher-2 is another database, which allows us to resolve packages to their URL and hash.
For details, see implementation.md.
After all python dependencies and their providers have been determined by the dependency resolver, mach-nix will generate a nix expression defining your python environment.
Individual python packages are either built by overriding an existing package definition from nixpkgs, or by creating the package from scratch via nixpkgs' buildPythonPackage. Which strategy is used depends on the provider of a package and if it is already packaged in nixpkgs.
Using nixpkgs as a base has the following benefits:
- Non-python Dependencies: Many python packages have non-python dependencies like various C libraries. Mach-nix can resolve those dependencies by taking the build inputs from python package definitions in nixpkgs.
- Special features: Some python packages can be built with special features, like for example SSE/AVX/FMA support in tensorflow. The nixpkgs versions of those python packages often include these features.
- Nix specific fixes: Some python packages might need some additional modification to work with nix. Those are already done in nixpkgs.
If a package is built by overriding nixpkgs, the following attributes are modified:
src: updated to the required versionname: modified to match the new versionbuildInputs: replaced with mach-nix determined python depspropagatedBuildInputs: non-python deps of old definition + mach-nix determined python depsdoCheck: set to false by default if not specified by userdoInstallCheck: set to false by default if not specified by user
Contributions to this project are welcome in the form of GitHub PRs. If you are planning to make any considerable changes, you should first present your plans in a GitHub issue so it can be discussed.
- Currently mach-nix does not provide any functionality which supports you in publishing python projects, like Poetry does for example.
![dependabot[bot] avatar](https://avatars.githubusercontent.com/in/29110?v=4 "dependabot[bot]")
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent