Why do we need another Haskell infrastructure for Nix?
Doesn't nixpkgs provide a sufficiently good Haskell infrastructure already?
Problems with the nixpkgs haskell infrastructure are covered in the following sections:
nixpkgs has quite good support for cross compilation, however the
Haskell infrastructure suffers from the fact that it heavily relies on
cabal2nix (as well as tools that depend on it
stack2nix) flattens the
.cabal file at conversion time to a
given os/arch/flags configuration. Thus to make cross compilation
cabal2nix you will have to generate a separate
expression for each configuration. This becomes a major maintenance
burden over time. Therefore the tooling that translates cabal files
into nix-expressions for use with Haskell.nix retains the full
conditional tree from the cabal file and exposes it to
addition it will also expose the
build-type value, which allows us
to cache the
Setup.hs for build-type simple and not have to rebuild
it every time.
We often rely on either package sets as provided by stackage or
computed by cabal.
nixpkgs provides its own curated package set
which might or might not work for the projects we work on.
stack2nix tries to solve this issue, here we go one step further and
provide the infrastructure to allow any form of package set.
Per component level control¶
The Haskell builder in
nixpkgs provides control over executables and
libraries, to build a specific executable only however is rather
tricky to do. This also leads to the cyclic dependencies issue.
The Haskell builder in
nixpkgs exposes packages at the
package level. If packages mutually depend on each other through tests
and libraries, this leads to cyclic dependencies that nix can't resolve. By
exposing the components to nix as separate derivations this will only
occur if you have mutually dependent components.
The Haskell builder in nixpkgs builds a package sequentially, first the library then the executables and finally the tests. It then executes the tests before the package is considered done. The upshot of this is that packages are only considered done if the test-suites passed. The downside is that if you have to compile multiple packages the likelihood of them failing is low, you have unnecessarily serialized your build. In a more aggressive setting libraries could start building as early as their dependent libraries are built. Of course they will have to be invalidated later should the test-suites of their dependencies fail, but this way we can make use of parallel building. In an ideal scenario this will reduce build times close to the optimum.
More logic in nix¶
cabal2nix tool has a resolver that resolves system dependencies
and licenses to values in
nixpkgs. This logic ends up being a simple
dictionary lookup and therefore can be a simple nix expression. This also
offloads some of the work the cabal to nix translation tool needs to
do into nix, and as such if changes are necessary (or needed to be
performed ad hoc) there is no need to rebuild the conversion tool and
subsequently mark every derived expression as out of date.
Finally, by treating Haskell.nix and nixpkgs as separate entities we can decouple the Haskell packages and infrastructure from the nixpkgs package set, and rely on it to provide us with system packages while staying up to date with Haskell packages from hackage while retaining a stable (or known to be good) nixpkgs revision.