So many preludes!

January 2, 2013

GravatarMichael Snoyman

I'm happy to announce new versions of basic-prelude, classy-prelude, and classy-prelude-conduit, as well as the addition of the new package classy-prelude-yesod. This is the first release of these packages which I consider stable enough for general use, and I encourage people to have a look. You can also check out the basic-prelude and classy-prelude repos on Github.

Since it's been a while since I discussed classy-prelude, let's start with a quick recap, beginning with the motivation: I think that the standard Prelude is lacking in a few ways:

  1. It encourages some bad practices (e.g., the partial head function).
  2. It promotes certain datatypes (e.g., String) over better alternatives (e.g., Text).
  3. Some commonly used functions are not exported (e.g., mapMaybe).
  4. Since it sticks to concrete types in many cases (usually lists), it uses up the common namespace for function names (e.g., length), thereby requiring programmers to use qualified imports on a regular basis.

I think the first point stands on its own: there's a basic question we need to ask ourselves about what we consider idiomatic Haskell code, and in my opinion, partial functions is not a part of it. While that's an important point to discuss, it's relatively straight-forward, so I won't be dwelling on it any further.

The other three points rest around a central philosophy I have: programmers are inherently lazy (in a good way), and will often choose the path of least resistance. To demonstrate, consider the difference between using String and Text for some simple concatenation:

-- String version
name = firstName ++ " " ++ lastName

-- Text version
import qualified Data.Text as T
name = firstName `T.append` " " `T.append` lastName

(Without OverloadedStrings, this would be even longer.) It's not that the second version is really that much worse than the first, it's just slightly less convenient. And due to that extra bit of work, Text gets used less often. You can see the same thing with Map versus associated lists, Vector and lists, and so on.

Note: As Herbert pointed out to me, with GHC 7.4 and up, you could just use the <> operator provided by Data.Monoid instance of `T.append`. So consider the case where you need to use some Prelude functions that require a String.

-- String version
main = putStrLn $ "Invalid name: " ++ name

-- Text version
import qualified Data.Text as T
import qualified Data.Text.IO as TIO
import Data.Monoid ((<>))
main = TIO.putStrLn $ "Invalid name: " <> name

By comparison, in classy-prelude this becomes:

import ClassyPrelude
main = putStrLn $ "Invalid name: " ++ name

If you think that my assessment so far doesn't warrant any changes to our tooling, turn back now, this blog post isn't for you. If you are interested in some kind of a solution to this issue, I have two options for you.

basic-prelude

Points 1-3 above can actually be solved pretty easily: just create a new prelude module that has better defaults. BasicPrelude, provided by the basic-prelude package, does just this. It exports more helper functions, avoids partial functions, and exports a bunch of the common datatypes, like ByteString and HashMap.

Another important premise of BasicPrelude is that it doesn't replace any existing types or type classes. It reuses the exact same Monad typeclass that is exported by Prelude. That means that code using BasicPrelude is completely compatible with "normal" Haskell code. Another way to put it is that BasicPrelude is a non-revolutionary approach to improving the Haskell programming experience.

basic-prelude is actually split up into two modules: BasicPrelude and CorePrelude. BasicPrelude simply re-exports everything provided by CorePrelude, and then adds in some missing components. CorePrelude is intended to be a foundation for the creation of other preludes. It was originally part of classy-prelude, but was then separated out by Dan Burton, who now maintains basic-prelude with me. CorePrelude tries to export components that would be usable by all Prelude replacements. For now, our simple barometer of this is "would both BasicPrelude and ClassyPrelude use this?"

BasicPrelude sticks to monomorphic functions for the most part, with a strong bias towards lists (just like standard Prelude). It doesn't really do much that's controversial, and should be a great approach to try out for people experimenting with an alternate prelude. And if you don't like something about it, you can either file an issue, or just create your own fork. Due to the design of basic-prelude, forking does not create incompatible code, so it's not a high-impact move.

classy-prelude

classy-prelude is the more radical prelude. As mentioned, it builds on top of CorePrelude, just like BasicPrelude does. The distinction, however, is that instead of providing monomorphic, list-biased functions, it creates a slew of typeclasses and provides polymorphic functions. Unlike many common typeclasses, these typeclasses are not intended to be used in a polymorphic context themselves, but rather to avoid the need to use qualified imports to disambiguate names. In other words, we're using typeclasses for namespacing purposes only.

(Despite this, we actually have a fairly thorough test suite covering the behavioral laws of these type classes. So you could theoretically write polymorphic code with classy-prelude, it's just not what I originally intended.)

This namespacing approach was fairly uncommon (perhaps classy-prelude was the first usage of it?) when I first started classy-prelude, and as a result I was unsure how well it would turn out in practice. At this point, I've been using classy-prelude for a number of my projects (both personal and at work), and the approach is certainly viable. I personally greatly prefer it to the non-classy approach, and will almost certainly be using it for the foreseeable future- likely until we get a better namespacing solution in GHC itself.

There are of course some downsides, some of which can be worked around:

  • Error messages become more confusing. I have no good solution to that right now. I don't think the messages are too daunting for experienced Haskellers, but I would not recommend classy-prelude to beginners.

  • In some cases it is impossible for the compiler to figure out which type you mean. For example, the following monomorphic code is completely unambiguous:

    import qualified Data.Map as Map
    
    foo :: Text -> Int
    foo name =
        Map.lookup name people
      where
        people = Map.singleton "Michael" 28

    However, the equivalent classy code is problematic:

    foo :: Text -> Int
    foo name =
        lookup name people
      where
        people = singleton "Michael" 28

    The problem is that both singleton and lookup are polymorphic, and are not used in the result, so there's no way to know which container to use. Fortunately, there's an easy workaround: the as* functions. In our case, we just replace the last line with:

        people = asMap $ singleton "Michael" 28

    Overall, the code is still shorter. As an added bonus, you can now simply switch asMap to asHashMap to swap which container structure you use.

  • In some cases, keeping the same name can go beyond the capabilities of the type system. For example, when working on classy-prelude-yesod, overloading insert for both its usage in Persistent and containers like Map proved to be a bit problematic, specifically when not using the return value. For example, the following code doesn't compile:

    _ <- runDB $ insert $ Person "Michael" 28

    The options in this case are either to not use the overloaded name and instead use a separately named function (in this case, insertDB), or to use a disambiguating helper function (voidKey) that fixes the types- similar to the asMap function described above. Those two solutions look like:

    voidKey $ runDB $ insert $ Person "Michael" 28
    runDB $ insertDB $ Person "Michael" 28

    I'm not sure yet how I feel about these two approaches, but it definitely stresses the point that we're using the typeclass system in an extreme manner.

classy-prelude-conduit and classy-prelude-yesod

These two packages build on top of classy-prelude to provide even more common functionality. The former provides conduit and xml-conduit functions, while the latter adds on top of that yesod, persistent, http-conduit, and a few other things. classy-prelude-yesod has not been as thoroughly exercised as the other packages discussed here, so you're more likely to run into issues with it.

Conclusion

These alternate preludes are not for everyone, but I think they offer a lot to certain audiences. If you want to try out a smaller move, I'd recommend BasicPrelude. If you want to be a bit more experimental, go for classy-prelude. If you decide to drop either one at any point, it should not be overly onerous to switch back to normal, monomorphic, qualified imports.

I'm definitely interested to hear people's experience with these packages. There are still lots of improvements to be made, more common functionality to be added, and documentation to be written. Now's a great time to get involved!

Comments

comments powered by Disqus

Archives