nixpkgs/doc/languages-frameworks/python.section.md
Jan Tojnar 6ecc641d08
doc: prepare for commonmark
We are still using Pandoc’s Markdown parser, which differs from CommonMark spec slightly.

Notably:
- Line breaks in lists behave differently.
- Admonitions do not support the simpler syntax https://github.com/jgm/commonmark-hs/issues/75
- The auto_identifiers uses a different algorithm – I made the previous ones explicit.
- Languages (classes) of code blocks cannot contain whitespace so we have to use “pycon” alias instead of Python “console” as GitHub’s linguist

While at it, I also fixed the following issues:
- ShellSesssion was used
- Removed some pointless docbook tags.
2021-06-07 06:34:59 +02:00

59 KiB
Raw Blame History

Python

User Guide

Using Python

Overview

Several versions of the Python interpreter are available on Nix, as well as a high amount of packages. The attribute python3 refers to the default interpreter, which is currently CPython 3.8. The attribute python refers to CPython 2.7 for backwards-compatibility. It is also possible to refer to specific versions, e.g. python38 refers to CPython 3.8, and pypy refers to the default PyPy interpreter.

Python is used a lot, and in different ways. This affects also how it is packaged. In the case of Python on Nix, an important distinction is made between whether the package is considered primarily an application, or whether it should be used as a library, i.e., of primary interest are the modules in site-packages that should be importable.

In the Nixpkgs tree Python applications can be found throughout, depending on what they do, and are called from the main package set. Python libraries, however, are in separate sets, with one set per interpreter version.

The interpreters have several common attributes. One of these attributes is pkgs, which is a package set of Python libraries for this specific interpreter. E.g., the toolz package corresponding to the default interpreter is python.pkgs.toolz, and the CPython 3.8 version is python38.pkgs.toolz. The main package set contains aliases to these package sets, e.g. pythonPackages refers to python.pkgs and python38Packages to python38.pkgs.

Installing Python and packages

The Nix and NixOS manuals explain how packages are generally installed. In the case of Python and Nix, it is important to make a distinction between whether the package is considered an application or a library.

Applications on Nix are typically installed into your user profile imperatively using nix-env -i, and on NixOS declaratively by adding the package name to environment.systemPackages in /etc/nixos/configuration.nix. Dependencies such as libraries are automatically installed and should not be installed explicitly.

The same goes for Python applications. Python applications can be installed in your profile, and will be wrapped to find their exact library dependencies, without impacting other applications or polluting your user environment.

But Python libraries you would like to use for development cannot be installed, at least not individually, because they won't be able to find each other resulting in import errors. Instead, it is possible to create an environment with python.buildEnv or python.withPackages where the interpreter and other executables are wrapped to be able to find each other and all of the modules.

In the following examples we will start by creating a simple, ad-hoc environment with a nix-shell that has numpy and toolz in Python 3.8; then we will create a re-usable environment in a single-file Python script; then we will create a full Python environment for development with this same environment.

Philosphically, this should be familiar to users who are used to a venv style of development: individual projects create their own Python environments without impacting the global environment or each other.

Ad-hoc temporary Python environment with nix-shell

The simplest way to start playing with the way nix wraps and sets up Python environments is with nix-shell at the cmdline. These environments create a temporary shell session with a Python and a precise list of packages (plus their runtime dependencies), with no other Python packages in the Python interpreter's scope.

To create a Python 3.8 session with numpy and toolz available, run:

$ nix-shell -p 'python38.withPackages(ps: with ps; [ numpy toolz ])'

By default nix-shell will start a bash session with this interpreter in our PATH, so if we then run:

[nix-shell:~/src/nixpkgs]$ python3
Python 3.8.1 (default, Dec 18 2019, 19:06:26)
[GCC 9.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy; import toolz

Note that no other modules are in scope, even if they were imperatively installed into our user environment as a dependency of a Python application:

>>> import requests
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ModuleNotFoundError: No module named 'requests'

We can add as many additional modules onto the nix-shell as we need, and we will still get 1 wrapped Python interpreter. We can start the interpreter directly like so:

$ nix-shell -p 'python38.withPackages(ps: with ps; [ numpy toolz requests ])' --run python3
these derivations will be built:
  /nix/store/xbdsrqrsfa1yva5s7pzsra8k08gxlbz1-python3-3.8.1-env.drv
building '/nix/store/xbdsrqrsfa1yva5s7pzsra8k08gxlbz1-python3-3.8.1-env.drv'...
created 277 symlinks in user environment
Python 3.8.1 (default, Dec 18 2019, 19:06:26)
[GCC 9.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import requests
>>>

Notice that this time it built a new Python environment, which now includes requests. Building an environment just creates wrapper scripts that expose the selected dependencies to the interpreter while re-using the actual modules. This means if any other env has installed requests or numpy in a different context, we don't need to recompile them -- we just recompile the wrapper script that sets up an interpreter pointing to them. This matters much more for "big" modules like pytorch or tensorflow.

Module names usually match their names on pypi.org, but you can use the Nixpkgs search website to find them as well (along with non-python packages).

At this point we can create throwaway experimental Python environments with arbitrary dependencies. This is a good way to get a feel for how the Python interpreter and dependencies work in Nix and NixOS, but to do some actual development, we'll want to make it a bit more persistent.

Running Python scripts and using nix-shell as shebang

Sometimes, we have a script whose header looks like this:

#!/usr/bin/env python3
import numpy as np
a = np.array([1,2])
b = np.array([3,4])
print(f"The dot product of {a} and {b} is: {np.dot(a, b)}")

Executing this script requires a python3 that has numpy. Using what we learned in the previous section, we could startup a shell and just run it like so:

$ nix-shell -p 'python38.withPackages(ps: with ps; [ numpy ])' --run 'python3 foo.py'
The dot product of [1 2] and [3 4] is: 11

But if we maintain the script ourselves, and if there are more dependencies, it may be nice to encode those dependencies in source to make the script re-usable without that bit of knowledge. That can be done by using nix-shell as a shebang, like so:

#!/usr/bin/env nix-shell
#!nix-shell -i python3 -p "python3.withPackages(ps: [ ps.numpy ])"
import numpy as np
a = np.array([1,2])
b = np.array([3,4])
print(f"The dot product of {a} and {b} is: {np.dot(a, b)}")

Then we simply execute it, without requiring any environment setup at all!

$ ./foo.py
The dot product of [1 2] and [3 4] is: 11

If the dependencies are not available on the host where foo.py is executed, it will build or download them from a Nix binary cache prior to starting up, prior that it is executed on a machine with a multi-user nix installation.

This provides a way to ship a self bootstrapping Python script, akin to a statically linked binary, where it can be run on any machine (provided nix is installed) without having to assume that numpy is installed globally on the system.

By default it is pulling the import checkout of Nixpkgs itself from our nix channel, which is nice as it cache aligns with our other package builds, but we can make it fully reproducible by pinning the nixpkgs import:

#!/usr/bin/env nix-shell
#!nix-shell -i python3 -p "python3.withPackages(ps: [ ps.numpy ])"
#!nix-shell -I nixpkgs=https://github.com/NixOS/nixpkgs/archive/d373d80b1207d52621961b16aa4a3438e4f98167.tar.gz
import numpy as np
a = np.array([1,2])
b = np.array([3,4])
print(f"The dot product of {a} and {b} is: {np.dot(a, b)}")

This will execute with the exact same versions of Python 3.8, numpy, and system dependencies a year from now as it does today, because it will always use exactly git commit d373d80b1207d52621961b16aa4a3438e4f98167 of Nixpkgs for all of the package versions.

This is also a great way to ensure the script executes identically on different servers.

Load environment from .nix expression

We've now seen how to create an ad-hoc temporary shell session, and how to create a single script with Python dependencies, but in the course of normal development we're usually working in an entire package repository.

As explained in the Nix manual, nix-shell can also load an expression from a .nix file. Say we want to have Python 3.8, numpy and toolz, like before, in an environment. We can add a shell.nix file describing our dependencies:

with import <nixpkgs> {};
(python38.withPackages (ps: [ps.numpy ps.toolz])).env

And then at the command line, just typing nix-shell produces the same environment as before. In a normal project, we'll likely have many more dependencies; this can provide a way for developers to share the environments with each other and with CI builders.

What's happening here?

  1. We begin with importing the Nix Packages collections. import <nixpkgs> imports the <nixpkgs> function, {} calls it and the with statement brings all attributes of nixpkgs in the local scope. These attributes form the main package set.
  2. Then we create a Python 3.8 environment with the withPackages function, as before.
  3. The withPackages function expects us to provide a function as an argument that takes the set of all Python packages and returns a list of packages to include in the environment. Here, we select the packages numpy and toolz from the package set.

To combine this with mkShell you can:

with import <nixpkgs> {};
let
  pythonEnv = python38.withPackages (ps: [
    ps.numpy
    ps.toolz
  ]);
in mkShell {
  packages = [
    pythonEnv

    black
    mypy

    libffi
    openssl
  ];
}

This will create a unified environment that has not just our Python interpreter and its Python dependencies, but also tools like black or mypy and libraries like libffi the openssl in scope. This is generic and can span any number of tools or languages across the Nixpkgs ecosystem.

Installing environments globally on the system

Up to now, we've been creating environments scoped to an ad-hoc shell session, or a single script, or a single project. This is generally advisable, as it avoids pollution across contexts.

However, sometimes we know we will often want a Python with some basic packages, and want this available without having to enter into a shell or build context. This can be useful to have things like vim/emacs editors and plugins or shell tools "just work" without having to set them up, or when running other software that expects packages to be installed globally.

To create your own custom environment, create a file in ~/.config/nixpkgs/overlays/ that looks like this:

# ~/.config/nixpkgs/overlays/myEnv.nix
self: super: {
  myEnv = super.buildEnv {
    name = "myEnv";
    paths = [
      # A Python 3 interpreter with some packages
      (self.python3.withPackages (
        ps: with ps; [
          pyflakes
          pytest
          python-language-server
        ]
      ))

      # Some other packages we'd like as part of this env
      self.mypy
      self.black
      self.ripgrep
      self.tmux
    ];
  };
}

You can then build and install this to your profile with:

nix-env -iA myEnv

One limitation of this is that you can only have 1 Python env installed globally, since they conflict on the python to load out of your PATH.

If you get a conflict or prefer to keep the setup clean, you can have nix-env atomically uninstall all other imperatively installed packages and replace your profile with just myEnv by using the --replace flag.

Environment defined in /etc/nixos/configuration.nix

For the sake of completeness, here's how to install the environment system-wide on NixOS.

{ # ...

  environment.systemPackages = with pkgs; [
    (python38.withPackages(ps: with ps; [ numpy toolz ]))
  ];
}

Developing with Python

Above, we were mostly just focused on use cases and what to do to get started creating working Python environments in nix.

Now that you know the basics to be up and running, it is time to take a step back and take a deeper look at how Python packages are packaged on Nix. Then, we will look at how you can use development mode with your code.

Python library packages in Nixpkgs

With Nix all packages are built by functions. The main function in Nix for building Python libraries is buildPythonPackage. Let's see how we can build the toolz package.

{ lib, buildPythonPackage, fetchPypi }:

buildPythonPackage rec {
  pname = "toolz";
  version = "0.10.0";

  src = fetchPypi {
    inherit pname version;
    sha256 = "08fdd5ef7c96480ad11c12d472de21acd32359996f69a5259299b540feba4560";
  };

  doCheck = false;

  meta = with lib; {
    homepage = "https://github.com/pytoolz/toolz";
    description = "List processing tools and functional utilities";
    license = licenses.bsd3;
    maintainers = with maintainers; [ fridh ];
  };
}

What happens here? The function buildPythonPackage is called and as argument it accepts a set. In this case the set is a recursive set, rec. One of the arguments is the name of the package, which consists of a basename (generally following the name on PyPi) and a version. Another argument, src specifies the source, which in this case is fetched from PyPI using the helper function fetchPypi. The argument doCheck is used to set whether tests should be run when building the package. Furthermore, we specify some (optional) meta information. The output of the function is a derivation.

An expression for toolz can be found in the Nixpkgs repository. As explained in the introduction of this Python section, a derivation of toolz is available for each interpreter version, e.g. python38.pkgs.toolz refers to the toolz derivation corresponding to the CPython 3.8 interpreter.

The above example works when you're directly working on pkgs/top-level/python-packages.nix in the Nixpkgs repository. Often though, you will want to test a Nix expression outside of the Nixpkgs tree.

The following expression creates a derivation for the toolz package, and adds it along with a numpy package to a Python environment.

with import <nixpkgs> {};

( let
    my_toolz = python38.pkgs.buildPythonPackage rec {
      pname = "toolz";
      version = "0.10.0";

      src = python38.pkgs.fetchPypi {
        inherit pname version;
        sha256 = "08fdd5ef7c96480ad11c12d472de21acd32359996f69a5259299b540feba4560";
      };

      doCheck = false;

      meta = {
        homepage = "https://github.com/pytoolz/toolz/";
        description = "List processing tools and functional utilities";
      };
    };

  in python38.withPackages (ps: [ps.numpy my_toolz])
).env

Executing nix-shell will result in an environment in which you can use Python 3.8 and the toolz package. As you can see we had to explicitly mention for which Python version we want to build a package.

So, what did we do here? Well, we took the Nix expression that we used earlier to build a Python environment, and said that we wanted to include our own version of toolz, named my_toolz. To introduce our own package in the scope of withPackages we used a let expression. You can see that we used ps.numpy to select numpy from the nixpkgs package set (ps). We did not take toolz from the Nixpkgs package set this time, but instead took our own version that we introduced with the let expression.

Handling dependencies

Our example, toolz, does not have any dependencies on other Python packages or system libraries. According to the manual, buildPythonPackage uses the arguments buildInputs and propagatedBuildInputs to specify dependencies. If something is exclusively a build-time dependency, then the dependency should be included in buildInputs, but if it is (also) a runtime dependency, then it should be added to propagatedBuildInputs. Test dependencies are considered build-time dependencies and passed to checkInputs.

The following example shows which arguments are given to buildPythonPackage in order to build datashape.

{ lib, buildPythonPackage, fetchPypi, numpy, multipledispatch, dateutil, pytest }:

buildPythonPackage rec {
  pname = "datashape";
  version = "0.4.7";

  src = fetchPypi {
    inherit pname version;
    sha256 = "14b2ef766d4c9652ab813182e866f493475e65e558bed0822e38bf07bba1a278";
  };

  checkInputs = [ pytest ];
  propagatedBuildInputs = [ numpy multipledispatch dateutil ];

  meta = with lib; {
    homepage = "https://github.com/ContinuumIO/datashape";
    description = "A data description language";
    license = licenses.bsd2;
    maintainers = with maintainers; [ fridh ];
  };
}

We can see several runtime dependencies, numpy, multipledispatch, and dateutil. Furthermore, we have one checkInputs, i.e. pytest. pytest is a test runner and is only used during the checkPhase and is therefore not added to propagatedBuildInputs.

In the previous case we had only dependencies on other Python packages to consider. Occasionally you have also system libraries to consider. E.g., lxml provides Python bindings to libxml2 and libxslt. These libraries are only required when building the bindings and are therefore added as buildInputs.

{ lib, pkgs, buildPythonPackage, fetchPypi }:

buildPythonPackage rec {
  pname = "lxml";
  version = "3.4.4";

  src = fetchPypi {
    inherit pname version;
    sha256 = "16a0fa97hym9ysdk3rmqz32xdjqmy4w34ld3rm3jf5viqjx65lxk";
  };

  buildInputs = [ pkgs.libxml2 pkgs.libxslt ];

  meta = with lib; {
    description = "Pythonic binding for the libxml2 and libxslt libraries";
    homepage = "https://lxml.de";
    license = licenses.bsd3;
    maintainers = with maintainers; [ sjourdois ];
  };
}

In this example lxml and Nix are able to work out exactly where the relevant files of the dependencies are. This is not always the case.

The example below shows bindings to The Fastest Fourier Transform in the West, commonly known as FFTW. On Nix we have separate packages of FFTW for the different types of floats ("single", "double", "long-double"). The bindings need all three types, and therefore we add all three as buildInputs. The bindings don't expect to find each of them in a different folder, and therefore we have to set LDFLAGS and CFLAGS.

{ lib, pkgs, buildPythonPackage, fetchPypi, numpy, scipy }:

buildPythonPackage rec {
  pname = "pyFFTW";
  version = "0.9.2";

  src = fetchPypi {
    inherit pname version;
    sha256 = "f6bbb6afa93085409ab24885a1a3cdb8909f095a142f4d49e346f2bd1b789074";
  };

  buildInputs = [ pkgs.fftw pkgs.fftwFloat pkgs.fftwLongDouble];

  propagatedBuildInputs = [ numpy scipy ];

  # Tests cannot import pyfftw. pyfftw works fine though.
  doCheck = false;

  preConfigure = ''
    export LDFLAGS="-L${pkgs.fftw.dev}/lib -L${pkgs.fftwFloat.out}/lib -L${pkgs.fftwLongDouble.out}/lib"
    export CFLAGS="-I${pkgs.fftw.dev}/include -I${pkgs.fftwFloat.dev}/include -I${pkgs.fftwLongDouble.dev}/include"
  '';

  meta = with lib; {
    description = "A pythonic wrapper around FFTW, the FFT library, presenting a unified interface for all the supported transforms";
    homepage = "http://hgomersall.github.com/pyFFTW";
    license = with licenses; [ bsd2 bsd3 ];
    maintainers = with maintainers; [ fridh ];
  };
}

Note also the line doCheck = false;, we explicitly disabled running the test-suite.

Testing Python Packages

It is highly encouraged to have testing as part of the package build. This helps to avoid situations where the package was able to build and install, but is not usable at runtime. Currently, all packages will use the test command provided by the setup.py (i.e. python setup.py test). However, this is currently deprecated https://github.com/pypa/setuptools/pull/1878 and your package should provide its own checkPhase.

NOTE: The checkPhase for python maps to the installCheckPhase on a normal derivation. This is due to many python packages not behaving well to the pre-installed version of the package. Version info, and natively compiled extensions generally only exist in the install directory, and thus can cause issues when a test suite asserts on that behavior.

NOTE: Tests should only be disabled if they don't agree with nix (e.g. external dependencies, network access, flakey tests), however, as many tests should be enabled as possible. Failing tests can still be a good indication that the package is not in a valid state.

Using pytest

Pytest is the most common test runner for python repositories. A trivial test run would be:

  checkInputs = [ pytest ];
  checkPhase = "pytest";

However, many repositories' test suites do not translate well to nix's build sandbox, and will generally need many tests to be disabled.

To filter tests using pytest, one can do the following:

  checkInputs = [ pytest ];
  # avoid tests which need additional data or touch network
  checkPhase = ''
    pytest tests/ --ignore=tests/integration -k 'not download and not update'
  '';

--ignore will tell pytest to ignore that file or directory from being collected as part of a test run. This is useful is a file uses a package which is not available in nixpkgs, thus skipping that test file is much easier than having to create a new package.

-k is used to define a predicate for test names. In this example, we are filtering out tests which contain download or update in their test case name. Only one -k argument is allowed, and thus a long predicate should be concatenated with “\” and wrapped to the next line.

NOTE: In pytest==6.0.1, the use of “\” to continue a line (e.g. -k 'not download \') has been removed, in this case, it's recommended to use pytestCheckHook.

Using pytestCheckHook

pytestCheckHook is a convenient hook which will substitute the setuptools test command for a checkPhase which runs pytest. This is also beneficial when a package may need many items disabled to run the test suite.

Using the example above, the analagous pytestCheckHook usage would be:

  checkInputs = [ pytestCheckHook ];

  # requires additional data
  pytestFlagsArray = [ "tests/" "--ignore=tests/integration" ];

  disabledTests = [
    # touches network
    "download"
    "update"
  ];

  disabledTestPaths = [
    "tests/test_failing.py"
  ];

This is expecially useful when tests need to be conditionallydisabled, for example:

  disabledTests = [
    # touches network
    "download"
    "update"
  ] ++ lib.optionals (pythonAtLeast "3.8") [
    # broken due to python3.8 async changes
    "async"
  ] ++ lib.optionals stdenv.isDarwin [
    # can fail when building with other packages
    "socket"
  ];

Trying to concatenate the related strings to disable tests in a regular checkPhase would be much harder to read. This also enables us to comment on why specific tests are disabled.

Using pythonImportsCheck

Although unit tests are highly prefered to validate correctness of a package, not all packages have test suites that can be ran easily, and some have none at all. To help ensure the package still works, pythonImportsCheck can attempt to import the listed modules.

  pythonImportsCheck = [ "requests" "urllib" ];

roughly translates to:

  postCheck = ''
    PYTHONPATH=$out/${python.sitePackages}:$PYTHONPATH
    python -c "import requests; import urllib"
  '';

However, this is done in it's own phase, and not dependent on whether doCheck = true;

This can also be useful in verifying that the package doesn't assume commonly present packages (e.g. setuptools)

Develop local package

As a Python developer you're likely aware of development mode (python setup.py develop); instead of installing the package this command creates a special link to the project code. That way, you can run updated code without having to reinstall after each and every change you make. Development mode is also available. Let's see how you can use it.

In the previous Nix expression the source was fetched from an url. We can also refer to a local source instead using src = ./path/to/source/tree;

If we create a shell.nix file which calls buildPythonPackage, and if src is a local source, and if the local source has a setup.py, then development mode is activated.

In the following example we create a simple environment that has a Python 3.8 version of our package in it, as well as its dependencies and other packages we like to have in the environment, all specified with propagatedBuildInputs. Indeed, we can just add any package we like to have in our environment to propagatedBuildInputs.

with import <nixpkgs> {};
with python38Packages;

buildPythonPackage rec {
  name = "mypackage";
  src = ./path/to/package/source;
  propagatedBuildInputs = [ pytest numpy pkgs.libsndfile ];
}

It is important to note that due to how development mode is implemented on Nix it is not possible to have multiple packages simultaneously in development mode.

Organising your packages

So far we discussed how you can use Python on Nix, and how you can develop with it. We've looked at how you write expressions to package Python packages, and we looked at how you can create environments in which specified packages are available.

At some point you'll likely have multiple packages which you would like to be able to use in different projects. In order to minimise unnecessary duplication we now look at how you can maintain a repository with your own packages. The important functions here are import and callPackage.

Including a derivation using callPackage

Earlier we created a Python environment using withPackages, and included the toolz package via a let expression. Let's split the package definition from the environment definition.

We first create a function that builds toolz in ~/path/to/toolz/release.nix

{ lib, buildPythonPackage }:

buildPythonPackage rec {
  pname = "toolz";
  version = "0.10.0";

  src = fetchPypi {
    inherit pname version;
    sha256 = "08fdd5ef7c96480ad11c12d472de21acd32359996f69a5259299b540feba4560";
  };

  meta = with lib; {
    homepage = "https://github.com/pytoolz/toolz/";
    description = "List processing tools and functional utilities";
    license = licenses.bsd3;
    maintainers = with maintainers; [ fridh ];
  };
}

It takes an argument buildPythonPackage. We now call this function using callPackage in the definition of our environment

with import <nixpkgs> {};

( let
    toolz = callPackage /path/to/toolz/release.nix {
      buildPythonPackage = python38Packages.buildPythonPackage;
    };
  in python38.withPackages (ps: [ ps.numpy toolz ])
).env

Important to remember is that the Python version for which the package is made depends on the python derivation that is passed to buildPythonPackage. Nix tries to automatically pass arguments when possible, which is why generally you don't explicitly define which python derivation should be used. In the above example we use buildPythonPackage that is part of the set python38Packages, and in this case the python38 interpreter is automatically used.

Reference

Interpreters

Versions 2.7, 3.6, 3.7, 3.8 and 3.9 of the CPython interpreter are available as respectively python27, python36, python37, python38 and python39. The aliases python2 and python3 correspond to respectively python27 and python39. The attribute python maps to python2. The PyPy interpreters compatible with Python 2.7 and 3 are available as pypy27 and pypy3, with aliases pypy2 mapping to pypy27 and pypy mapping to pypy2. The Nix expressions for the interpreters can be found in pkgs/development/interpreters/python.

All packages depending on any Python interpreter get appended out/{python.sitePackages} to $PYTHONPATH if such directory exists.

Missing tkinter module standard library

To reduce closure size the Tkinter/tkinter is available as a separate package, pythonPackages.tkinter.

Attributes on interpreters packages

Each interpreter has the following attributes:

  • libPrefix. Name of the folder in ${python}/lib/ for corresponding interpreter.
  • interpreter. Alias for ${python}/bin/${executable}.
  • buildEnv. Function to build python interpreter environments with extra packages bundled together. See section python.buildEnv function for usage and documentation.
  • withPackages. Simpler interface to buildEnv. See section python.withPackages function for usage and documentation.
  • sitePackages. Alias for lib/${libPrefix}/site-packages.
  • executable. Name of the interpreter executable, e.g. python3.8.
  • pkgs. Set of Python packages for that specific interpreter. The package set can be modified by overriding the interpreter and passing packageOverrides.

Optimizations

The Python interpreters are by default not build with optimizations enabled, because the builds are in that case not reproducible. To enable optimizations, override the interpreter of interest, e.g using

let
  pkgs = import ./. {};
  mypython = pkgs.python3.override {
    enableOptimizations = true;
    reproducibleBuild = false;
    self = mypython;
  };
in mypython

Building packages and applications

Python libraries and applications that use setuptools or distutils are typically built with respectively the buildPythonPackage and buildPythonApplication functions. These two functions also support installing a wheel.

All Python packages reside in pkgs/top-level/python-packages.nix and all applications elsewhere. In case a package is used as both a library and an application, then the package should be in pkgs/top-level/python-packages.nix since only those packages are made available for all interpreter versions. The preferred location for library expressions is in pkgs/development/python-modules. It is important that these packages are called from pkgs/top-level/python-packages.nix and not elsewhere, to guarantee the right version of the package is built.

Based on the packages defined in pkgs/top-level/python-packages.nix an attribute set is created for each available Python interpreter. The available sets are

  • pkgs.python27Packages
  • pkgs.python36Packages
  • pkgs.python37Packages
  • pkgs.python38Packages
  • pkgs.python39Packages
  • pkgs.pypyPackages

and the aliases

  • pkgs.python2Packages pointing to pkgs.python27Packages
  • pkgs.python3Packages pointing to pkgs.python38Packages
  • pkgs.pythonPackages pointing to pkgs.python2Packages

buildPythonPackage function

The buildPythonPackage function is implemented in pkgs/development/interpreters/python/mk-python-derivation using setup hooks.

The following is an example:

{ lib, buildPythonPackage, fetchPypi, hypothesis, setuptools-scm, attrs, py, setuptools, six, pluggy }:

buildPythonPackage rec {
  pname = "pytest";
  version = "3.3.1";

  src = fetchPypi {
    inherit pname version;
    sha256 = "cf8436dc59d8695346fcd3ab296de46425ecab00d64096cebe79fb51ecb2eb93";
  };

  postPatch = ''
    # don't test bash builtins
    rm testing/test_argcomplete.py
  '';

  checkInputs = [ hypothesis ];
  nativeBuildInputs = [ setuptools-scm ];
  propagatedBuildInputs = [ attrs py setuptools six pluggy ];

  meta = with lib; {
    maintainers = with maintainers; [ domenkozar lovek323 madjar lsix ];
    description = "Framework for writing tests";
  };
}

The buildPythonPackage mainly does four things:

  • In the buildPhase, it calls ${python.interpreter} setup.py bdist_wheel to build a wheel binary zipfile.
  • In the installPhase, it installs the wheel file using pip install *.whl.
  • In the postFixup phase, the wrapPythonPrograms bash function is called to wrap all programs in the $out/bin/* directory to include $PATH environment variable and add dependent libraries to script's sys.path.
  • In the installCheck phase, ${python.interpreter} setup.py test is ran.

By default tests are run because doCheck = true. Test dependencies, like e.g. the test runner, should be added to checkInputs.

By default meta.platforms is set to the same value as the interpreter unless overridden otherwise.

buildPythonPackage parameters

All parameters from stdenv.mkDerivation function are still supported. The following are specific to buildPythonPackage:

  • catchConflicts ? true: If true, abort package build if a package name appears more than once in dependency tree. Default is true.
  • disabled ? false: If true, package is not built for the particular Python interpreter version.
  • dontWrapPythonPrograms ? false: Skip wrapping of Python programs.
  • permitUserSite ? false: Skip setting the PYTHONNOUSERSITE environment variable in wrapped programs.
  • format ? "setuptools": Format of the source. Valid options are "setuptools", "pyproject", "flit", "wheel", and "other". "setuptools" is for when the source has a setup.py and setuptools is used to build a wheel, flit, in case flit should be used to build a wheel, and wheel in case a wheel is provided. Use other when a custom buildPhase and/or installPhase is needed.
  • makeWrapperArgs ? []: A list of strings. Arguments to be passed to makeWrapper, which wraps generated binaries. By default, the arguments to makeWrapper set PATH and PYTHONPATH environment variables before calling the binary. Additional arguments here can allow a developer to set environment variables which will be available when the binary is run. For example, makeWrapperArgs = ["--set FOO BAR" "--set BAZ QUX"].
  • namePrefix: Prepends text to ${name} parameter. In case of libraries, this defaults to "python3.8-" for Python 3.8, etc., and in case of applications to "".
  • pipInstallFlags ? []: A list of strings. Arguments to be passed to pip install. To pass options to python setup.py install, use --install-option. E.g., pipInstallFlags=["--install-option='--cpp_implementation'"].
  • pythonPath ? []: List of packages to be added into $PYTHONPATH. Packages in pythonPath are not propagated (contrary to propagatedBuildInputs).
  • preShellHook: Hook to execute commands before shellHook.
  • postShellHook: Hook to execute commands after shellHook.
  • removeBinByteCode ? true: Remove bytecode from /bin. Bytecode is only created when the filenames end with .py.
  • setupPyGlobalFlags ? []: List of flags passed to setup.py command.
  • setupPyBuildFlags ? []: List of flags passed to setup.py build_ext command.

The stdenv.mkDerivation function accepts various parameters for describing build inputs (see "Specifying dependencies"). The following are of special interest for Python packages, either because these are primarily used, or because their behaviour is different:

  • nativeBuildInputs ? []: Build-time only dependencies. Typically executables as well as the items listed in setup_requires.
  • buildInputs ? []: Build and/or run-time dependencies that need to be compiled for the host machine. Typically non-Python libraries which are being linked.
  • checkInputs ? []: Dependencies needed for running the checkPhase. These are added to nativeBuildInputs when doCheck = true. Items listed in tests_require go here.
  • propagatedBuildInputs ? []: Aside from propagating dependencies, buildPythonPackage also injects code into and wraps executables with the paths included in this list. Items listed in install_requires go here.
Overriding Python packages

The buildPythonPackage function has a overridePythonAttrs method that can be used to override the package. In the following example we create an environment where we have the blaze package using an older version of pandas. We override first the Python interpreter and pass packageOverrides which contains the overrides for packages in the package set.

with import <nixpkgs> {};

(let
  python = let
    packageOverrides = self: super: {
      pandas = super.pandas.overridePythonAttrs(old: rec {
        version = "0.19.1";
        src =  super.fetchPypi {
          pname = "pandas";
          inherit version;
          sha256 = "08blshqj9zj1wyjhhw3kl2vas75vhhicvv72flvf1z3jvapgw295";
        };
      });
    };
  in pkgs.python3.override {inherit packageOverrides; self = python;};

in python.withPackages(ps: [ps.blaze])).env

buildPythonApplication function

The buildPythonApplication function is practically the same as buildPythonPackage. The main purpose of this function is to build a Python package where one is interested only in the executables, and not importable modules. For that reason, when adding this package to a python.buildEnv, the modules won't be made available.

Another difference is that buildPythonPackage by default prefixes the names of the packages with the version of the interpreter. Because this is irrelevant for applications, the prefix is omitted.

When packaging a Python application with buildPythonApplication, it should be called with callPackage and passed python or pythonPackages (possibly specifying an interpreter version), like this:

{ lib, python3Packages }:

python3Packages.buildPythonApplication rec {
  pname = "luigi";
  version = "2.7.9";

  src = python3Packages.fetchPypi {
    inherit pname version;
    sha256 = "035w8gqql36zlan0xjrzz9j4lh9hs0qrsgnbyw07qs7lnkvbdv9x";
  };

  propagatedBuildInputs = with python3Packages; [ tornado_4 python-daemon ];

  meta = with lib; {
    ...
  };
}

This is then added to all-packages.nix just as any other application would be.

luigi = callPackage ../applications/networking/cluster/luigi { };

Since the package is an application, a consumer doesn't need to care about Python versions or modules, which is why they don't go in pythonPackages.

toPythonApplication function

A distinction is made between applications and libraries, however, sometimes a package is used as both. In this case the package is added as a library to python-packages.nix and as an application to all-packages.nix. To reduce duplication the toPythonApplication can be used to convert a library to an application.

The Nix expression shall use buildPythonPackage and be called from python-packages.nix. A reference shall be created from all-packages.nix to the attribute in python-packages.nix, and the toPythonApplication shall be applied to the reference:

youtube-dl = with pythonPackages; toPythonApplication youtube-dl;

toPythonModule function

In some cases, such as bindings, a package is created using stdenv.mkDerivation and added as attribute in all-packages.nix. The Python bindings should be made available from python-packages.nix. The toPythonModule function takes a derivation and makes certain Python-specific modifications.

opencv = toPythonModule (pkgs.opencv.override {
  enablePython = true;
  pythonPackages = self;
});

Do pay attention to passing in the right Python version!

python.buildEnv function

Python environments can be created using the low-level pkgs.buildEnv function. This example shows how to create an environment that has the Pyramid Web Framework. Saving the following as default.nix

with import <nixpkgs> {};

python.buildEnv.override {
  extraLibs = [ pythonPackages.pyramid ];
  ignoreCollisions = true;
}

and running nix-build will create

/nix/store/cf1xhjwzmdki7fasgr4kz6di72ykicl5-python-2.7.8-env

with wrapped binaries in bin/.

You can also use the env attribute to create local environments with needed packages installed. This is somewhat comparable to virtualenv. For example, running nix-shell with the following shell.nix

with import <nixpkgs> {};

(python3.buildEnv.override {
  extraLibs = with python3Packages; [ numpy requests ];
}).env

will drop you into a shell where Python will have the specified packages in its path.

python.buildEnv arguments
  • extraLibs: List of packages installed inside the environment.
  • postBuild: Shell command executed after the build of environment.
  • ignoreCollisions: Ignore file collisions inside the environment (default is false).
  • permitUserSite: Skip setting the PYTHONNOUSERSITE environment variable in wrapped binaries in the environment.

python.withPackages function

The python.withPackages function provides a simpler interface to the python.buildEnv functionality. It takes a function as an argument that is passed the set of python packages and returns the list of the packages to be included in the environment. Using the withPackages function, the previous example for the Pyramid Web Framework environment can be written like this:

with import <nixpkgs> {};

python.withPackages (ps: [ps.pyramid])

withPackages passes the correct package set for the specific interpreter version as an argument to the function. In the above example, ps equals pythonPackages. But you can also easily switch to using python3:

with import <nixpkgs> {};

python3.withPackages (ps: [ps.pyramid])

Now, ps is set to python3Packages, matching the version of the interpreter.

As python.withPackages simply uses python.buildEnv under the hood, it also supports the env attribute. The shell.nix file from the previous section can thus be also written like this:

with import <nixpkgs> {};

(python38.withPackages (ps: [ps.numpy ps.requests])).env

In contrast to python.buildEnv, python.withPackages does not support the more advanced options such as ignoreCollisions = true or postBuild. If you need them, you have to use python.buildEnv.

Python 2 namespace packages may provide __init__.py that collide. In that case python.buildEnv should be used with ignoreCollisions = true.

Setup hooks

The following are setup hooks specifically for Python packages. Most of these are used in buildPythonPackage.

  • eggUnpackhook to move an egg to the correct folder so it can be installed with the eggInstallHook
  • eggBuildHook to skip building for eggs.
  • eggInstallHook to install eggs.
  • flitBuildHook to build a wheel using flit.
  • pipBuildHook to build a wheel using pip and PEP 517. Note a build system (e.g. setuptools or flit) should still be added as nativeBuildInput.
  • pipInstallHook to install wheels.
  • pytestCheckHook to run tests with pytest. See example usage.
  • pythonCatchConflictsHook to check whether a Python package is not already existing.
  • pythonImportsCheckHook to check whether importing the listed modules works.
  • pythonRemoveBinBytecode to remove bytecode from the /bin folder.
  • setuptoolsBuildHook to build a wheel using setuptools.
  • setuptoolsCheckHook to run tests with python setup.py test.
  • venvShellHook to source a Python 3 venv at the venvDir location. A venv is created if it does not yet exist. postVenvCreation can be used to to run commands only after venv is first created.
  • wheelUnpackHook to move a wheel to the correct folder so it can be installed with the pipInstallHook.

Development mode

Development or editable mode is supported. To develop Python packages buildPythonPackage has additional logic inside shellPhase to run pip install -e . --prefix $TMPDIR/for the package.

Warning: shellPhase is executed only if setup.py exists.

Given a default.nix:

with import <nixpkgs> {};

pythonPackages.buildPythonPackage {
  name = "myproject";
  buildInputs = with pythonPackages; [ pyramid ];

  src = ./.;
}

Running nix-shell with no arguments should give you the environment in which the package would be built with nix-build.

Shortcut to setup environments with C headers/libraries and Python packages:

nix-shell -p pythonPackages.pyramid zlib libjpeg git

Note: There is a boolean value lib.inNixShell set to true if nix-shell is invoked.

Tools

Packages inside nixpkgs are written by hand. However many tools exist in community to help save time. No tool is preferred at the moment.

Deterministic builds

The Python interpreters are now built deterministically. Minor modifications had to be made to the interpreters in order to generate deterministic bytecode. This has security implications and is relevant for those using Python in a nix-shell.

When the environment variable DETERMINISTIC_BUILD is set, all bytecode will have timestamp 1. The buildPythonPackage function sets DETERMINISTIC_BUILD=1 and PYTHONHASHSEED=0. Both are also exported in nix-shell.

Automatic tests

It is recommended to test packages as part of the build process. Source distributions (sdist) often include test files, but not always.

By default the command python setup.py test is run as part of the checkPhase, but often it is necessary to pass a custom checkPhase. An example of such a situation is when py.test is used.

Common issues

  • Non-working tests can often be deselected. By default buildPythonPackage runs python setup.py test. Most Python modules follows the standard test protocol where the pytest runner can be used instead. py.test supports a -k parameter to ignore test methods or classes:

    buildPythonPackage {
      # ...
      # assumes the tests are located in tests
      checkInputs = [ pytest ];
      checkPhase = ''
        py.test -k 'not function_name and not other_function' tests
      '';
    }
    
  • Tests that attempt to access $HOME can be fixed by using the following work-around before running tests (e.g. preCheck): export HOME=$(mktemp -d)

FAQ

How to solve circular dependencies?

Consider the packages A and B that depend on each other. When packaging B, a solution is to override package A not to depend on B as an input. The same should also be done when packaging A.

How to override a Python package?

We can override the interpreter and pass packageOverrides. In the following example we rename the pandas package and build it.

with import <nixpkgs> {};

(let
  python = let
    packageOverrides = self: super: {
      pandas = super.pandas.overridePythonAttrs(old: {name="foo";});
    };
  in pkgs.python38.override {inherit packageOverrides;};

in python.withPackages(ps: [ps.pandas])).env

Using nix-build on this expression will build an environment that contains the package pandas but with the new name foo.

All packages in the package set will use the renamed package. A typical use case is to switch to another version of a certain package. For example, in the Nixpkgs repository we have multiple versions of django and scipy. In the following example we use a different version of scipy and create an environment that uses it. All packages in the Python package set will now use the updated scipy version.

with import <nixpkgs> {};

( let
    packageOverrides = self: super: {
      scipy = super.scipy_0_17;
    };
  in (pkgs.python38.override {inherit packageOverrides;}).withPackages (ps: [ps.blaze])
).env

The requested package blaze depends on pandas which itself depends on scipy.

If you want the whole of Nixpkgs to use your modifications, then you can use overlays as explained in this manual. In the following example we build a inkscape using a different version of numpy.

let
  pkgs = import <nixpkgs> {};
  newpkgs = import pkgs.path { overlays = [ (self: super: {
    python38 = let
      packageOverrides = python-self: python-super: {
        numpy = python-super.numpy_1_18;
      };
    in super.python38.override {inherit packageOverrides;};
  } ) ]; };
in newpkgs.inkscape

python setup.py bdist_wheel cannot create .whl

Executing python setup.py bdist_wheel in a nix-shell fails with

ValueError: ZIP does not support timestamps before 1980

This is because files from the Nix store (which have a timestamp of the UNIX epoch of January 1, 1970) are included in the .ZIP, but .ZIP archives follow the DOS convention of counting timestamps from 1980.

The command bdist_wheel reads the SOURCE_DATE_EPOCH environment variable, which nix-shell sets to 1. Unsetting this variable or giving it a value corresponding to 1980 or later enables building wheels.

Use 1980 as timestamp:

nix-shell --run "SOURCE_DATE_EPOCH=315532800 python3 setup.py bdist_wheel"

or the current time:

nix-shell --run "SOURCE_DATE_EPOCH=$(date +%s) python3 setup.py bdist_wheel"

or unset SOURCE_DATE_EPOCH:

nix-shell --run "unset SOURCE_DATE_EPOCH; python3 setup.py bdist_wheel"

install_data / data_files problems

If you get the following error:

could not create '/nix/store/6l1bvljpy8gazlsw2aw9skwwp4pmvyxw-python-2.7.8/etc':
Permission denied

This is a known bug in setuptools. Setuptools install_data does not respect --prefix. An example of such package using the feature is pkgs/tools/X11/xpra/default.nix.

As workaround install it as an extra preInstall step:

${python.interpreter} setup.py install_data --install-dir=$out --root=$out
sed -i '/ = data\_files/d' setup.py

Rationale of non-existent global site-packages

On most operating systems a global site-packages is maintained. This however becomes problematic if you want to run multiple Python versions or have multiple versions of certain libraries for your projects. Generally, you would solve such issues by creating virtual environments using virtualenv.

On Nix each package has an isolated dependency tree which, in the case of Python, guarantees the right versions of the interpreter and libraries or packages are available. There is therefore no need to maintain a global site-packages.

If you want to create a Python environment for development, then the recommended method is to use nix-shell, either with or without the python.buildEnv function.

How to consume Python modules using pip in a virtual environment like I am used to on other Operating Systems?

While this approach is not very idiomatic from Nix perspective, it can still be useful when dealing with pre-existing projects or in situations where it's not feasible or desired to write derivations for all required dependencies.

This is an example of a default.nix for a nix-shell, which allows to consume a virtual environment created by venv, and install Python modules through pip the traditional way.

Create this default.nix file, together with a requirements.txt and simply execute nix-shell.

with import <nixpkgs> { };

let
  pythonPackages = python3Packages;
in pkgs.mkShell rec {
  name = "impurePythonEnv";
  venvDir = "./.venv";
  buildInputs = [
    # A Python interpreter including the 'venv' module is required to bootstrap
    # the environment.
    pythonPackages.python

    # This execute some shell code to initialize a venv in $venvDir before
    # dropping into the shell
    pythonPackages.venvShellHook

    # Those are dependencies that we would like to use from nixpkgs, which will
    # add them to PYTHONPATH and thus make them accessible from within the venv.
    pythonPackages.numpy
    pythonPackages.requests

    # In this particular example, in order to compile any binary extensions they may
    # require, the Python modules listed in the hypothetical requirements.txt need
    # the following packages to be installed locally:
    taglib
    openssl
    git
    libxml2
    libxslt
    libzip
    zlib
  ];

  # Run this command, only after creating the virtual environment
  postVenvCreation = ''
    unset SOURCE_DATE_EPOCH
    pip install -r requirements.txt
  '';

  # Now we can execute any commands within the virtual environment.
  # This is optional and can be left out to run pip manually.
  postShellHook = ''
    # allow pip to install wheels
    unset SOURCE_DATE_EPOCH
  '';

}

In case the supplied venvShellHook is insufficient, or when Python 2 support is needed, you can define your own shell hook and adapt to your needs like in the following example:

with import <nixpkgs> { };

let
  venvDir = "./.venv";
  pythonPackages = python3Packages;
in pkgs.mkShell rec {
  name = "impurePythonEnv";
  buildInputs = [
    pythonPackages.python
    # Needed when using python 2.7
    # pythonPackages.virtualenv
    # ...
  ];

  # This is very close to how venvShellHook is implemented, but
  # adapted to use 'virtualenv'
  shellHook = ''
    SOURCE_DATE_EPOCH=$(date +%s)

    if [ -d "${venvDir}" ]; then
      echo "Skipping venv creation, '${venvDir}' already exists"
    else
      echo "Creating new venv environment in path: '${venvDir}'"
      # Note that the module venv was only introduced in python 3, so for 2.7
      # this needs to be replaced with a call to virtualenv
      ${pythonPackages.python.interpreter} -m venv "${venvDir}"
    fi

    # Under some circumstances it might be necessary to add your virtual
    # environment to PYTHONPATH, which you can do here too;
    # PYTHONPATH=$PWD/${venvDir}/${pythonPackages.python.sitePackages}/:$PYTHONPATH

    source "${venvDir}/bin/activate"

    # As in the previous example, this is optional.
    pip install -r requirements.txt
  '';
}

Note that the pip install is an imperative action. So every time nix-shell is executed it will attempt to download the Python modules listed in requirements.txt. However these will be cached locally within the virtualenv folder and not downloaded again.

How to override a Python package from configuration.nix?

If you need to change a package's attribute(s) from configuration.nix you could do:

  nixpkgs.config.packageOverrides = super: {
    python = super.python.override {
      packageOverrides = python-self: python-super: {
        twisted = python-super.twisted.overrideAttrs (oldAttrs: {
          src = super.fetchPipy {
            pname = "twisted";
            version = "19.10.0";
            sha256 = "7394ba7f272ae722a74f3d969dcf599bc4ef093bc392038748a490f1724a515d";
            extension = "tar.bz2";
          };
        });
      };
    };
  };

pythonPackages.twisted is now globally overridden. All packages and also all NixOS services that reference twisted (such as services.buildbot-worker) now use the new definition. Note that python-super refers to the old package set and python-self to the new, overridden version.

To modify only a Python package set instead of a whole Python derivation, use this snippet:

  myPythonPackages = pythonPackages.override {
    overrides = self: super: {
      twisted = ...;
    };
  }

How to override a Python package using overlays?

Use the following overlay template:

self: super: {
  python = super.python.override {
    packageOverrides = python-self: python-super: {
      twisted = python-super.twisted.overrideAttrs (oldAttrs: {
        src = super.fetchPypi {
          pname = "twisted";
          version = "19.10.0";
          sha256 = "7394ba7f272ae722a74f3d969dcf599bc4ef093bc392038748a490f1724a515d";
          extension = "tar.bz2";
        };
      });
    };
  };
}

How to use Intels MKL with numpy and scipy?

MKL can be configured using an overlay. See the section "Using overlays to configure alternatives".

What inputs do setup_requires, install_requires and tests_require map to?

In a setup.py or setup.cfg it is common to declare dependencies:

  • setup_requires corresponds to nativeBuildInputs
  • install_requires corresponds to propagatedBuildInputs
  • tests_require corresponds to checkInputs

Contributing

Contributing guidelines

The following rules are desired to be respected:

  • Python libraries are called from python-packages.nix and packaged with buildPythonPackage. The expression of a library should be in pkgs/development/python-modules/<name>/default.nix.
  • Python applications live outside of python-packages.nix and are packaged with buildPythonApplication.
  • Make sure libraries build for all Python interpreters.
  • By default we enable tests. Make sure the tests are found and, in the case of libraries, are passing for all interpreters. If certain tests fail they can be disabled individually. Try to avoid disabling the tests altogether. In any case, when you disable tests, leave a comment explaining why.
  • Commit names of Python libraries should reflect that they are Python libraries, so write for example pythonPackages.numpy: 1.11 -> 1.12.
  • Attribute names in python-packages.nix as well as pnames should match the library's name on PyPI, but be normalized according to PEP 0503. This means that characters should be converted to lowercase and . and _ should be replaced by a single - (foo-bar-baz instead of Foo__Bar.baz). If necessary, pname has to be given a different value within fetchPypi.
  • Attribute names in python-packages.nix should be sorted alphanumerically to avoid merge conflicts and ease locating attributes.