Automatically add build-system.requires to nativeBuildInputs

[GuillaumeDesforges]

EDIT: original title was "flit_core backend missing"

Describe the issue

In entrypoints-0.4 the backend is flit_core.buildapi which is specified in requires in build-system.
https://github.com/takluyver/entrypoints/blob/ebdf2d8edc9921427ea07688851999796093c240/pyproject.toml#L2-L3

Trying to build entrypoints in a poetry2nix mkPoetryEnv fails on build with

ModuleNotFoundError: No module named 'flit_core'

Additional context

Entering nix shell I can check that flit_core is not available (not in pip list)
It's weird because it seems to me that poetry2nix reads the backend fields

poetry2nix/lib.nix

Line 161 in 69cc39a

getBuildSystemPkgs =

• default.nix/shell.nix/flake.nix

{
  inputs.flake-utils.url = "github:numtide/flake-utils";
  inputs.poetry2nix.url = "github:nix-community/poetry2nix";

  outputs = { self, nixpkgs, flake-utils, poetry2nix, ... }:
    flake-utils.lib.eachDefaultSystem (system:
      let
        pkgs = import nixpkgs {
          inherit system;
          overlay = [ poetry2nix.overlay ];
        };
      in
      {
        devShell =
          pkgs.poetry2nix.mkPoetryEnv {
            projectDir = ./.;
          };
      });
}

• pyproject.toml

[ tool.poetry ]
name = "poetry-nix"
version = "0.1.0"
description = ""
authors = ["Guillaume Desforges <[email protected]>"]

[tool.poetry.dependencies]
python = "^3.8"
entrypoints = "^0.4"

[tool.poetry.dev-dependencies]

[build-system]
requires = ["poetry-core>=1.0.0"]
build-backend = "poetry.core.masonry.api"

`poetry.lock`

[[package]]
name = "entrypoints"
version = "0.4"
description = "Discover and load entry points from installed packages."
category = "main"
optional = false
python-versions = ">=3.6"

[metadata]
lock-version = "1.1"
python-versions = "^3.8"
content-hash = "5d84993b5abc54589527347f1f6dcea0edde667a81ae537884ae1585c42b951a"

[metadata.files]
entrypoints = [
    {file = "entrypoints-0.4-py3-none-any.whl", hash = "sha256:f174b5ff827504fd3cd97cc3f8649f3693f51538c7e4bdf3ef002c8429d42f9f"},
    {file = "entrypoints-0.4.tar.gz", hash = "sha256:b706eddaa9218a19ebcd67b56818f05bb27589b1ca9e8d797b74affad4ccacd4"},
]

[flokli]

I just ran into this while trying to add structlog to a poetry2nix-packaged environment.

[flokli]

I read some code a bit, and added some traces.

It seems in general, there's all tooling in place to parse build-system.build-backend, and to add these dependencies.

However, it seems buildSystemPkgs isn't added to buildInputs (due to the ++ lib.optional isDirectory buildSystemPkgs in mk-poetry-dep.nix).

I don't quite understand the bigger picture yet, why this only applies to isDirectory, not isDirectory || isGit || isUrl like in other places. It also seems source is null for some reason.

@adisbladis, any idea?

[flokli]

I'm able to workaround this by manually bringing in a flit-core, via the overrides mechanism:

poetry2nix.mkPoetryApplication {
  projectDir = ./.;
  python = python310;
  overrides = poetry2nix.overrides.withDefaults (self: super: {
    # …
    # workaround https://github.com/nix-community/poetry2nix/issues/568
    structlog = super.structlog.overridePythonAttrs (old: {
      buildInputs = old.buildInputs or [ ] ++ [ python310.pkgs.flit-core ];
    });
  });
}

So it seems this indeed just a matter of not automatically bringing in the build system packages for dependencies.

[GuillaumeDesforges]

Yes, I'm using this workaround as well.

I've also seen that the build system dependencies are hard coded for now for a subset of packages.
https://github.com/nix-community/poetry2nix/blob/master/overrides/build-systems.json

It would be nice to parse the requires field in pyproject when available and add as nativeBuildInput.

[flokli]

Yeah, but the tooling should be there to just derive this from pyproject.toml. @GuillaumeDesforges do you want to rename this issue to reflect this?

[adisbladis]

The root cause of this issue is upstream python-poetry/poetry#2789.

[GuillaumeDesforges]

Since poetry does not write the build-system.requires to the lockfile, poetry2nix can't really get that information indeed.

[flokli]

Huuh. Can we link to the upstream issue somewhere in the poetry2nix docs, and provide some guidance on how we seem to be using the overrides mechanism to manually keep track of these dependencies?

[yajo]

Maybe could we include https://github.com/DavHau/pypi-deps-db and fetch the desired build system from there? Just like mach-nix does.

[andersk]

The upstream issue python-poetry/poetry#2789 was fixed, but the fix uses ephemeral build environments, and the information needed to reconstruct them is not recorded in the lock file. Perhaps another upstream issue should be opened explaining why it should be recorded?

[yajo]

I think upstream understands why it's important, but not how to implement it, due to a technical limitation they explained in python-poetry/poetry#5401 (comment).

Just keep that in mind, but in any case I agree to open a new issue and talk with upstream to see what solution could we get to. If solved upstream, it'd have obvious benefits upstream too, so I think they'll be open to it if there's a way to do it.

[nixos-discourse]

This issue has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/why-is-it-so-hard-to-use-a-python-package/19200/11

[FRidh]

I suggest doing the pinning of the build systems here in poetry2nix repo and let it update via actions periodically. It won't consider the version constraints that way, but it gets it working for probably 99% of the cases and is still reproducible.

Note though you will probably need to resolve them together, to ensure they can cooperate when needed. I think it is rare this is going to cause issues, but you don't know. This could be done with poetry itself.

[charmoniumQ]

The upstream issue python-poetry/poetry#2789 was fixed, but the fix uses ephemeral build environments, and the information needed to reconstruct them is not recorded in the lock file. Perhaps another upstream issue should be opened explaining why it should be recorded?

See python-poetry/poetry#6154

[colonelpanic8]

Looks like python-poetry/poetry#7975 was merged which may potentially fix this?

Has anyone tried overriding poetry with a git version to see if this fixes things?

[GuillaumeDesforges]

This example with entrypoints is not reproduced, probably because of 33db1f3

If anyone would like this issue to progress, we need another new reproducible example that makes sense.

[brendon-boldt]

This pyproject.toml fails for the same reason I believe (urllib3 needs hatchling which it can't find). I'd test out the most recent version of poetry in poetry2nix, but I couldn't figure out how to write that override.

[tool.poetry]
name = "test"
version = "0.0.0"
description = ""
authors = []

[tool.poetry.dependencies]
python = "3.10"
urllib3 = "2.0.4"

[tool.poetry.dev-dependencies]

[build-system]
requires = ["poetry-core>=1.0.0"]
build-backend = "poetry.core.masonry.api"

[dmytrokyrychuk]

@brendon-boldt I think your problem was resolved by 33db1f3, but that version of poetry2nix is not yet in nixpkgs. I tried your pyproject.toml with the latest poetry2nix flake, and urllib3 was installed correctly. See https://github.com/dmytrokyrychuk/nix-urllib3-test

[Turakar]

I think that the PyPI cmake package is an example of this:

pyproject.toml:

[tool.poetry]
...

[tool.poetry.dependencies]
python = "^3.11,<3.12"
cmake = "^3.27.7"


[build-system]
requires = ["poetry-core"]
build-backend = "poetry.core.masonry.api"

flake.nix:

{
  description = "Application packaged using poetry2nix";

  inputs = {
    flake-utils.url = "github:numtide/flake-utils";
    nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";
    poetry2nix = {
      url = "github:nix-community/poetry2nix";
      inputs.nixpkgs.follows = "nixpkgs";
    };
  };

  outputs = { self, nixpkgs, flake-utils, poetry2nix }:
    flake-utils.lib.eachDefaultSystem (system:
      let
        # see https://github.com/nix-community/poetry2nix/tree/master#api for more functions and examples.
        pkgs = nixpkgs.legacyPackages.${system};
        inherit (poetry2nix.lib.mkPoetry2Nix { inherit pkgs; }) mkPoetryApplication;
      in
      {
        packages = {
          myapp = mkPoetryApplication { projectDir = self; };
          default = self.packages.${system}.myapp ;
        };

        devShells.default = pkgs.mkShell {
          inputsFrom = [ self.packages.${system}.myapp ];
          packages = [ pkgs.poetry ];
        };
      });
}

Error:

$ nix develop
warning: Git tree '<path>' is dirty
error: builder for '/nix/store/bn6y3qp4mbxybgbbsz1azzka3d1fqpkw-python3.11-cmake-3.27.7.drv' failed with exit code 2;
       last 10 log lines:
       >   File "<frozen importlib._bootstrap>", line 1204, in _gcd_import
       >   File "<frozen importlib._bootstrap>", line 1176, in _find_and_load
       >   File "<frozen importlib._bootstrap>", line 1126, in _find_and_load_unlocked
       >   File "<frozen importlib._bootstrap>", line 241, in _call_with_frames_removed
       >   File "<frozen importlib._bootstrap>", line 1204, in _gcd_import
       >   File "<frozen importlib._bootstrap>", line 1176, in _find_and_load
       >   File "<frozen importlib._bootstrap>", line 1140, in _find_and_load_unlocked
       > ModuleNotFoundError: No module named 'setuptools'
       > 
       > 
       For full logs, run 'nix log /nix/store/bn6y3qp4mbxybgbbsz1azzka3d1fqpkw-python3.11-cmake-3.27.7.drv'.
error: 1 dependencies of derivation '/nix/store/djms9q6r0nl51c29d54kci9dibihzklx-nix-shell-env.drv' failed to build

[adisbladis]

See python-poetry/poetry#8752

[GuillaumeDesforges]

poetry is not going to cripple its performance just to add data to the lockfile that it does not need

IMO that makes sense, and it is best to figure out a way that does not rely on poetry changing its behavior.

It seems to me that most packages use only a very limited set of packages for build-system.requires.
Could poetry2nix read pyproject.toml and use a mapping from these names in build-system.requires to packages?

E.g.

{ ... }:
let
  # from nixpkgs
  pythonPackages = ...;
  pyprojectToml = ...;
  depToDrv = {
    # For example
    "poetry-core" = pythonPackages.poetry-core;
  };
  buildRequires = map (dep: ) pyprojectToml.build-system.requires;
in
  # add `buildRequires` to the Python build

[adisbladis]

It seems to me that most packages use only a very limited set of packages for build-system.requires.
Could poetry2nix read pyproject.toml and use a mapping from these names in build-system.requires to packages?

That would be IFD, so not possible.

[charmoniumQ]

If I understand the issue correctly, poetry2nix, being a Nix program, can only access the internet if we statically know a hash of what we expect as the response. Much of the time, poetry.lock has the necessary hashes, so poetry2nix can go download transitive (build- or run-time) dependencies of the current project.

However, Poetry does not put build-time dependencies of the transitive (build- or run-time) dependencies into poetry.lock, because Poetry does not necessarily build-from-source (sdist). In fact it prefers to use pre-built wheels for performance reasons. So it would have to do extra work (slow everyone down) to find out and lock the build-time dependencies.

PyPI exposes the hash of the source through an HTTP API, which is how Poetry puts a hash of the source of every dependency in the Poetry.lock even when it builds from source. However, Nix cannot choose its dependencies dynamically; it can't download something to discover the build dependencies; it has to know them before it starts to download anything. Therefore, this doesn't help us that much.

---

Maybe we should frame our request to Poetry as "please provide a poetry --build-from-source". Does Poetry already know how to build from source if no wheels are available for the platform? Then this shouldn't be too hard to implement. Once it's building from source, it won't be that hard to augment the lockfile (although the spec has to change to include an optional build requirements, if available). This strategy would not hinder any users who don't explicitly enable this option.

[andersk]

If the reason it would be slow for Poetry to add build-system information to poetry.lock is that this information isn’t available via the PyPI HTTP API, that sounds like an argument for making it available via the PyPI HTTP API.

Maybe we should frame our request to Poetry as "please provide a poetry --build-from-source".

That’s already there as the installer.no-binary setting. But it comes with this note:

“As with all configurations described here, this is a user specific configuration. This means that this is not taken into consideration when a lockfile is generated or dependencies are resolved. This is applied only when selecting which distribution for dependency should be installed into a Poetry managed environment.”

[charmoniumQ]

Are there any Python package managers that make a better lockfile? Does PDM do any better?

[SemMulder]

Are there any Python package managers that make a better lockfile? Does PDM do any better?

Apparently, it is quite feasible to roll your own by using the Poetry solver as a library.

See here how conda-lock does it:
https://github.com/conda/conda-lock/blob/4fc441f2f33b9570917d2453d4eadf3cdc1d95f4/conda_lock/pypi_solver.py

That wouldn't be an ideal situation, due to the maintenance burden. But it might be a solution.

[yajo]

How about a poetry2nix.lock file? We could have a simple script here that generates that. Then, add it to your project closure and move on. Of course, to be locked impurely, just like poetry does.

[GuillaumeDesforges]

A good thing about poetry2nix was the ability to plug it on any poetry repository, including those which we don't have write permissions. Requiring a file specific to poetry2nix in the source code repos would make poetr2nix much less useful.

[yajo]

I didn't say you need to put that file in the source repository... Since that lock file is only relevant for poetry2nix (and only if you build from sources, not from wheels), then it makes sense that it resides in the same repo as the instructions to build with nix. Example:

{ poetry2nix, python310, fetchgitArgs }:
poetry2nix.mkPoetryApplication rec {
  python = python310;
  pyproject = src + "/pyproject.toml";
  poetrylock = src + "/poetry.lock";
  poetry2nixlock = ./poetry2nix.lock;
  src = builtins.fetchgit fetchgitArgs;
}