nmattia
Type Nix through haskell
Most of snack's flakiness comes from the fact that the logic is written in Nix - which is untyped and doesn't really have test frameworks.
An idea I played with a while ago was to wrap Nix with haskell types: https://github.com/nmattia/nix-reloaded
It provides a nice way to write Nix by lifting all expressions to typed haskell values, e.g.
https://github.com/nmattia/nix-reloaded/blob/82489bd9d58e21cc4611b2c00c757a859f5fe70a/Test.hs#L16-L21
test1 :: IO ()
test1 = do
x <- withNixRepl $ \nixRepl -> do
seven <- bindVar nixRepl (3 + 4)
eval nixRepl $ Pkgs.map (+ seven) (toNix [1 .. 10])
x @=? ((+ 7) <$> [1 .. 10] :: [Int])
an expression like builtins.map would be lifted to a haskell value of type (Var a -> Var b) -> Var [a] -> Var [b] which can then be used and applied like any other haskell function.
What do you think about moving more of the logic to Haskell directly in addition to this? For example hpack.nix?
One of the goals of snack is to allow the user to run nix-build and everything should work. If we push some of the logic to haskell that means that, somehow, the nix build needs to first call something like snack dump-build-procedure > foo.nix and then do some import from derivation. If we manage to get everything nice and typed in nix-reloaded then we can do snack dump-nix-files > lib.nix and just upload lib.nix on GitHub, and everyone else can just use that.
Ah, I see! The build process should not even depend on snack itself, only nix. So this is basically goal 1 mentioned in the readme, but more explicit.
And I guess the output of snack dump-nix-files should not depend on the project at all too, correct? (EDIT: well except on the choice of which config files to use)
FWIW there's an existing https://github.com/regnat/ptyx project that attempts to add a type system to Nix.
Last time I talked to @regnat he mentioned it wasn't looking really good. However @jwiegley mentioned there's a rudimentary type system in https://github.com/haskell-nix/hnix, so that could work.