Haskell: Difference between revisions

From NixOS Wiki
imported>Andreabedini
Mention new features in cabal 2.0
imported>Srid
Add haskell-flake
Line 201: Line 201:


The [https://github.com/input-output-hk/haskell.nix haskell.nix] project allows you to have maximum flexibility (to create your own package set or use in teams with diverse people, some of them using stack, other using cabal, other using nix…). But this comes at the price of additional complexity.
The [https://github.com/input-output-hk/haskell.nix haskell.nix] project allows you to have maximum flexibility (to create your own package set or use in teams with diverse people, some of them using stack, other using cabal, other using nix…). But this comes at the price of additional complexity.
=== Using haskell-flake (flake-parts) ===
[https://haskell.flake.page/ haskell-flake] aims to simplify writing Nix for Haskell development through use of flake-parts module system. It uses `developPackage` and `shellFor` under the hood while exposing friendly module options API.


== Overrides ==
== Overrides ==

Revision as of 22:51, 5 March 2023

FAQ and resources

How to develop with Haskell and Nix

There are multiples ways to develop in Haskell on Nix depending on the simplicity of the project and on whether one want to benefit from the reproducibility offered by nix or not. Below is an image to help you in your choice:

Note that in the following, haskellPackages is a synonym of haskell.packages.ghcXYZ where XYZ is the current default version of GHC in nixpkgs. However you can use a different version by replacing haskellPackages with the wanted package, for instance use haskell.compiler.ghc884 to use GHC 8.8.4. You can get the full list of available GHC versions using:

$ nix-env -f "<nixpkgs>" -qaP -A haskell.compiler
haskell.compiler.ghc8107                 ghc-8.10.7
haskell.compiler.ghc884                  ghc-8.8.4
haskell.compiler.ghc902                  ghc-9.0.2
haskell.compiler.ghc924                  ghc-9.2.4
haskell.compiler.ghcHEAD                 ghc-9.3.20220406
haskell.compiler.ghc942                  ghc-9.4.2
…

Scripting

For simple scripts, you can directly use nix-shell to get a redistributable Haskell script that you can run on any Nix system with ./my-script.hs:

#!/usr/bin/env nix-shell
#!nix-shell --pure -i runghc -p "haskellPackages.ghcWithPackages (pkgs: [ pkgs.turtle ])"

main = do
  # do stuff
  putStrLn "Hello world from a distributable Haskell script!"

Read below if some packages are broken.

Directly using cabal (no nix caching/reproducibility)

Note that cabal is the basic Haskell tool used to configure builds and is internally used by all the Haskell's packaging methods (including stack and nix). If one does not care about the reproducibility/caching offered by nix, it is always possible to use cabal like in a normal system:

$  nix-shell -p "haskellPackages.ghcWithPackages (pkgs: with pkgs; [ cabal-install ])"
$ cabal init
…
$ cabal run
Up to date
Hello, Haskell!

Notes:

  • some packages may need additional libraries/programs, notably zlib, you should be able to add them as additional programs in the nix-shell option
  • since Cabal 2.0, cabal has acquired caching similar to nix (but not as powerful) and reproducibility (via the cabal.project file and the index-state option). See [1] for more information.

Using Stack (no nix caching)

Similarly you can use stack that let you find the appropriate version of the libraries for you if you do not want the caching offered by nix (stack will build all the dependencies):

$ nix-shell -p "haskellPackages.ghcWithPackages (pkgs: with pkgs; [ stack ])"
$ stack new my-project
$ cd my-project
$ stack build
$ stack exec my-project-exe

You can also use the features offered by stack to enable nix integration in order to use nix to install the non-haskell dependencies. You can read more here.

If you want to package your program using stack in nix, you can actually use haskell.lib.buildStackProject that is a wrapper around stdenv.mkDerivation that will call stack build for you… However because stack needs to download stuff you need to disable the sandbox using nix-build --option sandbox false. For instance if you want to compile a stack project that needs R, zeromq and zlib you can put the following into default.nix:

with (import <nixpkgs> { });
haskell.lib.buildStackProject {
  name = "HaskellR";
  buildInputs = [ R zeromq zlib ];
}

Using developPackage (use the nix packages set for haskell)

You can use also nix in place of stack to keep track of the dependencies in a reproducible way (note that while stack uses a solver to find a working set of dependencies, nix uses a fixed set of packages). Additionally you can benefit from the caching system offered by Nix. To that end, first create a cabal repository (nix also uses cabal internally):

$ nix-shell -p "haskellPackages.ghcWithPackages (pkgs: with pkgs; [ cabal-install ])" --run "cabal init"

And create a file default.nix containing:

let
  pkgs = import <nixpkgs> { }; # pin the channel to ensure reproducibility!
in
pkgs.haskellPackages.developPackage {
  root = ./.;
}

(You can find a list of options and documentation for developPackage in pkgs/development/haskell-modules/make-package-set.nix, note that it is a wrapper around callCabal2nixWithOptions with some additional functions to setup a development shell.)

Then you can build and run the program using:

$ nix-build
$ ./result/bin/yourprogram 

or run a nix-shell to use the standard development tools provided by cabal:

$ nix-build
$ ./result/bin/yourprogram 

Nix will automatically read the build-depends field in the *.cabal file to get the name of the dependencies and use the haskell packages provided in the configured package set provided by nix. Note that some of the packages present in the nix repository are broken (for instance because a package requires an older version of a library while nix only provides a recent version). For this reason it may be necessary to override some packages present in the nix package set as described below using the overrides and source-overrides attribute. Note that the source-overrides attribute can also turn out to be useful to load local libraries:

let
  pkgs = import <nixpkgs> { }; # pin the channel to ensure reproducibility!
in
pkgs.haskellPackages.developPackage {
  root = ./.;
  source-overrides = {
    mylibrary = ./mylibrary;
  };
}

However as I understand I guess that you will not be able to enter the shell before `mylibrary` fully compiles… hence the need for `shellFor` to work simultaneously on multiple projects.

Note that you may want to add tools needed either at compile time or a library at run time. For that, you can use the modifier field that is an arbitrary function to apply to the final haskell package (in particular you can apply the overrideCabal that we saw above). Notably, you can add nativeBuildInputs using pkgs.haskell.lib.addBuildTools and buildInputs using pkgs.haskell.lib.addExtraLibraries (for those of you that are curious to see how they are used in the final derivation, see here):

let
  pkgs = import <nixpkgs> { }; # pin the channel to ensure reproducibility!
in
pkgs.haskellPackages.developPackage {
  root = ./.;
  modifier = drv:
    pkgs.haskell.lib.addBuildTools drv (with pkgs.haskellPackages;
      [ cabal-install
        ghcid
      ]);
}


You can also get more details in this tutorial or in pkgs/development/haskell-modules/make-package-set.nix.

Using shellFor (multiple packages)

shellFor is similar to developPackage but (slightly) more complicated to also allow you to develop multiples packages at the same time (similar to cabal.project). Note that contrary to developPackage I don't think that shellFor can output a derivation.

The idea is to first extend/override the set of haskell packages in order to add your projects as additional haskell packages (for instance using haskellPackages.extend and packageSourceOverrides that just need a the path of the project to compile it), and then to use `haskellPackages.shellFor {packages= p: [p.myproject1 p.myproject2]}` to create a shell with all wanted packages.

For instance you can define your various projects in subfolders ./frontend and ./backend (you can use cabal init to create the content in each folder), then create a file cabal.project containing:

packages:
  frontend/
  backend/

Finally create a file shell.nix containing:

with import <nixpkgs> {};
# We add our packages to the haskell package set
(haskellPackages.extend (haskell.lib.compose.packageSourceOverrides {
  frontend = ./frontend;
  backend = ./backend;
}))
# We call on this set shellFor to drop us into a shell containing the dependencies of frontend and backend:
  .shellFor {
    packages = p: [p.frontend p.backend];
    withHoogle = true;
    buildInputs = [ pkgs.python pkgs.cabal-install ];
  }

then you can use cabal to develop incrementally your projects using for instance:

$ nix-shell
$ cabal new-build all

If you want to be able to compile a project non-incrementally with nix-build (say the backend in the above example) you can put in default.nix:

with import <nixpkgs> {};
# We add our packages to the haskell package set
(haskellPackages.extend (haskell.lib.compose.packageSourceOverrides {
  frontend = ./frontend;
  backend = ./backend;
})).backend

or if you want to create a single derivation file, you can use if pkgs.lib.inNixShell then … else … to output the shell when we start a shell and the packages when we want to build them. You can find here an example.

Using haskell.nix (for complex projects)

The haskell.nix project allows you to have maximum flexibility (to create your own package set or use in teams with diverse people, some of them using stack, other using cabal, other using nix…). But this comes at the price of additional complexity.

Using haskell-flake (flake-parts)

haskell-flake aims to simplify writing Nix for Haskell development through use of flake-parts module system. It uses `developPackage` and `shellFor` under the hood while exposing friendly module options API.

Overrides

Since nixpkgs tries to maintain a single package set (based on the package set of stackage, while the remaining packages are picked from the latest version on Hackage) instead of using a solver to meet all version constraints for a specific project, it turns out that sometimes packages are broken (they can also be broken for various other reasons). However, you may be able to unbreak this package yourself.

The first thing to check to try to unbreak a package is to check which GHC version is compatible with the package you want to use. You are maybe using a too old version… or too new. You can change the version of ghc using haskell.packages.ghcXYZ in place of haskellPackages as explained above.

Then, you will surely need to change some packages. If you are using developPackage as explained above you can either use a normal override that will be described below or a simpler source override to override only the source as:

pkgs.haskellPackages.developPackage {
  root = ./.;
  source-overrides = {
    # Let's say the GHC haskellPackages uses 1.6.0.0 and your test suite is incompatible with >= 1.6.0.0
    HUnit = "1.5.0.0";
  };
};

You can provide to source-overrides either:

  • a version number (it will be forwarded internally to callHackage, note that you do not need to specify any hash as nix is using a package all-cabal-hashes that contains all the cabal hashes, see callHackageDirect below if your package is not yet in all-cabal-hashes)
  • a path (that will be forwarded internally to callCabal2nix), you can use the usual fetchers like fetchurl or fetchFromGitHub to generate that path, or a local path if you want to use a local library.

You can also use the more powerful override system to change any property of the derivation. This works for instance with developPackage:

pkgs.haskellPackages.developPackage {
  root = ./.;
  overrides = self: super: { # self is the new package set, super is the old package set
    random = pkgs.haskell.lib.overrideCabal super.random {
      version = "1.1";
      sha256 = "sha256-txikEFfiWjpx32k6sP4iY9SS51lnmzwv6m6jOxcdOlo="; # Use an empty string before knowing the hash
      doCheck = false;
    };
  };
};

but you can also use overrides with ghcWithPackages. This example will for instance create a nix-shell where the library quipper is available:

{ pkgs ? import <nixpkgs> {} }:
with pkgs;
pkgs.mkShell {
  buildInputs =
    let
      # Quipper does not work with GHC 7.10 or 8.10. The versions currently supported are GHC 8.0, 8.2, 8.4, 8.6, and 8.8.
      myHaskell = pkgs.haskell.packages.ghc884.override {
        overrides = self: super: {
          # fixedprec needs random 1.1 or below
          random = pkgs.haskell.lib.overrideCabal super.random {
            version = "1.1";
            sha256 = "sha256-txikEFfiWjpx32k6sP4iY9SS51lnmzwv6m6jOxcdOlo=";
          };
          fixedprec = haskell.lib.markUnbroken super.fixedprec;
          quipper = haskell.lib.markUnbroken super.quipper;
        };
      };
    in
      [
        (myHaskell.ghcWithPackages (hpkgs: [
          hpkgs.quipper
        ]))
      ];
}


Note that overrideCabal takes as input the old package and the new attributes of the new package and outputs the new package. To see the full list of parameters that can be overridden, you can refer to this file.

Because some operations are very common, there exists some functions that call overrideCabal for you. For instance if you only want to disable checks and test suits for a package you can do mypackage = pkgs.haskell.lib.dontCheck super.mypackage and the above code also shows how to mark a package as unbroken. These functions are listed and documented in pkgs/development/haskell-modules/lib/compose.nix.

You can also use callHackageDirect (source and documentation here, you can see that it is a wrapper around callCabal2nix) to create a package using the hackage repository:

myHaskell = pkgs.haskellPackages.override {
  overrides = self: super: {
    mypackage = self.callHackageDirect {
      pkg = "mypackage";
      ver = "0.1.2.3";
      sha256 = ""; # The first time it will give you an error, replace the "" with the hash given in the error
    } {};
  };
};

Finally, if your package is not in hackage, you can simply use callCabal2nix, or the more advanced callCabal2nixWithOptions:

mypackage = self.callCabal2nix "mypackage" /path/to/package/or/fetcher {};

This can be useful also when your package needs some libraries.

Limitations

When using the cabal2nix tool, Nix does not pull a cabal package by respecting the constraint specified in the cabal file (see example). Issue is discussed here. You should be using `callCabal2nix` anyway.

IFD and Haskell

callCabal2nix, which is implicitly used for building Haskell projects, uses IFD.[2][3]. This means that since IFD is disabled by default in certain nix commands,[4] the following commands will be broken for Haskell projects whose flake output specifies multiple system attributes:

  • nix flake show
  • nix flake check