ClassyPrelude: The good, the bad, and the ugly

August 3, 2012

GravatarBy Michael Snoyman

First a message from our sponsors: Yesod 1.1 is fully tested and ready for release. I'm holding off until the beginning of next week, since I hate making major releases on a Friday (if there are any bugs, I won't be around on Saturday to fix them). If you're really eager to check it out, you can get the latest and greatest from Github.


I promised to write this blog post a week or two ago, but never got around to it. Dan Burton reminded me about it yesterday... so here's paying off my debt :).

So far, most of the discussion around classy-prelude has focused around the "classy" bits. As Felipe mentioned a few weeks ago, there's a lot more going on than just the classy parts. So I'd like to break open the discussion more and explain what exactly is going on.

Before we start, I think I need to give a bit of an explanation of my development strategy, as I think some people don't understand my approach. When attacking a problem for the first time, I throw in the kitchen sink. Many people prefer to take a more iterative approach, adding features one step at a time after a lot of careful testing. And in many cases, I think that's the only valid approach.

But for a project like classy-prelude, I'm interested in being much more experimental. That means I'm happy to throw in some half-baked ideas to see if they work. Don't mistake that as me saying I fully endorse and recommend this approach. On the contrary, as I hope this post makes clear, I'm highly skeptical of some the choices I made. But I'm also a firm believer that we won't know if the choices were bad unless we try them out.

So without further ado, let's break down the classy-prelude, starting with the "obvious good" category and moving further into the unknown. (And I think it goes without saying that statements like "obviously good" are subjective.)

Win: Full compatibility with standard libraries

I saw the contrary claim made many times in various discussions, so let's put it to rest. Unlike other prelude replacements, classy-prelude does not try to fix flaws in the base package. Functor is not a superclass of Monad, fail still exists, and Num is unchanged. That's not to say that I think that the base package did everything correctly (I don't), but rather, by sticking to the standard typeclasses, classy-prelude code should work with any existing library without a hitch.

From a user perspective: you should be able to use classy-prelude in one module, ditch it in another, depend on some packages that use it and others that don't, without any problem. In fact, I've done exactly that in some of my code. If you use classy-prelude in your library, your library users should not be affected beyond the fact that they have to install classy-prelude.

As an aside, I'd like to see some of those bigger changes make it into base ultimately, and I think an alternate prelude is a great way to play around with the ideas. But such an alternate prelude wouldn't be much use for real-world use, which is why I haven't taken that route with classy-prelude.

Win: Don't export partial functions

I really wish Prelude didn't export head, tail, and a bunch of other partial functions. So classy-prelude doesn't. More generally, I've been taking a whitelist approach to which Prelude functions are exported. If I've left out one of your favorite functions, it's likely because I just didn't get around to it yet, not due to any hatred of it. If there's some total function in Prelude that you'd like added, just send a pull request.

Win: Export commonly used functions from other modules

I prefer using <$> to fmap in most circumstances. I like to use first and &&&, and a huge number of my modules import mapMaybe. All of these functions/operators have names that are well recognized in the community, I believe are unambiguous, and are very generally useful. Since the main purpose of classy-prelude is to save you from extra keystrokes, let's just export them all by default.

The selection here is clearly going to be pretty opinionated. And the list isn't really that exhaustive right now (again, pull requests welcome). But while we can argue over exactly which functions should be exported, I think this is a pretty solid advantage.

Win: Export datatypes

How many modules have you written with lines looking like:

import Data.ByteString (ByteString)

And how many times do you just import Data.ByteString qualified and use S.ByteString? How often is ByteString imported from Data.ByteString.Lazy instead? And do you make the alias for the module S or B?

These are all thoroughly uninteresting questions. Typing out the code is boring. And trying to remember which convention is used in the particular module you're working on is just a pain. So in classy-prelude, we have a simple convention: export the datatype. If there's a lazy variant, prefix it with L. So we end up with ByteString, LByteString, Text, LText, Map, HashMap, and so on.

There's nothing earth-shattering here. It just removes an annoyance.

Probably a win: Generalize using existing typeclasses

I like Monoid a lot. I think it's a great typeclass. And I think it's a shame that in Prelude, concat only works for lists and not arbitrary Monoids. Similarly, I wish that ++ was just mappend. So in classy-prelude, I've done just that.

This is a bit controversial, because it can make error messages a bit more confusing. But let me reinforce something that I probably didn't clarify enough: classy-prelude is not intended for beginners. I don't think that a Haskell tutorial should be using it for examples. classy-prelude is intended for Haskellers who are already battle-tested with GHC's error messages, and won't shy away from some more general types.

One change I didn't make, but was considering, was using . and id from Control.Category. I'm not sure how much of a practical benefit that would provide people, while making error messages likely much more complex. But if people want that change, let's discuss it.

More debatable win: recommend better datatypes

I've gone on record many times saying that I dislike type FilePath = [Char]. system-filepath is (yet another) wonderful package by John Millikin, and I use it extensively in my personal code. (It's also making its way more solidly into the Yesod ecosystem, though we frankly don't do that much filesystem access in Yesod.)

In classy-prelude, I export system-filepath's FilePath type. This is likely the most "breaking change" we have, though it doesn't really prevent you from continuing to use [Char] for all your filepaths. classy-prelude also exports a bunch of the functions and operators for manipulating filepaths, like basename and </>.

Again, this is an opinionated decision. But I think moving over to system-filepath could be a huge win for the Haskell community in general, and regardless of classy-prelude's future, I hope it catches on.

More controversial: create a bunch of new typeclasses

Now we get to the fun stuff. classy-prelude defines a bunch of new typeclasses for functions which are commonly imported qualified. I mentioned it previously, but I don't think my point got across, so I'll say it again. The purpose here has nothing to do with equational reasoning. This is purely a technique for name overloading, nothing more.

Let's expand on that a bit. One of the nice things about typeclasses is that they are usually associated with a set of laws that define their behavior. This lets us write code against a typeclass and know, based on the laws, how it will behave. We don't need to know if we're dealing with a list of a Set, we know that mappend empty foo is the same as foo.

That's not my purpose here. The purpose is purely about name overloading. I haven't stated a semantics for what lookup needs to do, because the answer is it depends on the container. This means that classy-prelude is not decreasing the cognitive load of the programmer in any meaningful way. If there are semantic differences between the insert functions for a HashMap and a Map, you still need to know about them. (To my knowledge, no such difference exists.)

So let's say it again: the only purpose for this technique is to allow name overloading.

Based on this, I think it's fair to claim that classy-prelude doesn't make code more buggy due to its lack of associated laws. The libraries you're using already don't conform to a set of laws. classy-prelude is just making it easier to use them. If you write type-generic code and try to use it for multiple container types (not a practice I recommend), it's possible that you'll get buggy code due to semantic differences. But the same applies if you write a function using Data.Set, copy-paste it, and then replace Data.Set with Data.HashSet.

All that said, I'm not saying that a set of associated laws isn't a good thing. I'd love to add them. And if classy-prelude continues, I'm sure we will add them. But for the initial proof-of-concept experimental release, I don't think it was worth the investment of time.

One final idea here is to leverage some existing collections of typeclasses (Edward Kmett's reducers package in particular) instead of defining our own typeclasses. Again, for the proof-of-concept, that idea was premature, but going forward I'm hoping to look into it. There are obstacles we'd have to overcome (like ensuring good performance), but I don't think anything is insurmountable.

Very controversial: make those typeclasses work with conduit

I was surprised not to see as much of a discussion about this point as the previous one. In my mind, this was by far the most controversial choice in classy-prelude. It's fair to say that I put it in more out of curiosity if it would work than anything else.

Let's take filter as an example. A relatively simple typeclass to model it would look like:

class CanFilter container element | container -> element where
    filter :: (element -> bool) -> container -> container

This would work for lists, ByteString, Set, and so on. However, here's the (simplified) signature of filter from Data.Conduit.List:

filter :: (i -> Bool) -> Conduit i m i

This doesn't fit in with our CanFilter class above. CanFilter has a function that takes two arguments, whereas in conduit filter has just one argument.

So here's the trick. We recognize that the first argument is the same (a predicate), and just leverage currying. Recognizing that CanFilter's filter can be viewed as a function of one argument returning a function, we split it up into two typeclasses:

class CanFilter result element where
    filter :: (element -> bool) -> result
class CanFilterFunc container element | container -> element where
    filterFunc :: (element -> bool) -> container -> container

And then we define a single instance of CanFilter that uses CanFilterFunc for all cases where we return a function:

instance (container ~ container', CanFilterFunc container element)
  => CanFilter (container -> container') element where
    filter = filterFunc

(If you're curious about the equality constraint, see this excellent Stack Overflow answer.)

We can then separately define an instance for conduit, which does not overlap at all, and allows us to reuse the name filter in significantly different use cases. Frankly, I think it's pretty cool that we have this kind of flexibility in the type class system. But from a practical standpoint, I'm not sure it's really a great trade-off:

  • Since the two filter concepts work fairly differently, it may be more confusing than anything else to lump them together.

  • Error messages get significantly more confusing.

  • I don't think that conduit usage in this sense is prevalent enough to warrant the costs.

In other words, if there's something I'd want to cut out first from classy-prelude, it's the gymnastics which it pulls to accomodate conduit instances. It was definitely a fun part of the experiment, and I'm glad to have tested it. If anyone has an opinion either way on this, let me know.

Moving forward

There are a number of minor bugs in the typeclass definitions in classy-prelude, requiring more explicit type signatures than should be necessary in some cases. Those kinds of things are easily fixed. If people find specific examples, please bring them to my attention.

Dan Burton started a project called ModularPrelude which takes a different approach to the namespace issue, replicating first class modules via the record system. I think it's definitely an interesting approach, and think it can coexist very well with classy-prelude. But it's not enough for my taste: I prefer the relative terseness of typeclass-based code.

However, putting that project together with my points in this post, I think it's important to note that there is a very large part of classy-prelude which has nothing to do with typeclasses. Dan called this BasicPrelude.hs, and I think it makes a lot of sense to provide that separately (though at least for now, I'll keep it in the same package for convenience).

The idea is simple: I think a lot of people agree with me that some major aspects of this prelude are no-brainers, and would like to use them. People are understably more wary of the experimental bits. (As I've said before, so am I: I won't allow classy-prelude to be used in yesod in its current state.) So let's create a stable basis, and encourage experimentation. As I mentioned at the beginning, since we have full compatibility with base, there's no real concern of fragmentation.

Once I remove the conduit aspects of the library, I think it will open it up for much better analysis of typeclass laws. It might be possible to completely drop a number of typeclasses and use the reducers versions instead. For others, more experimentation might be necessary.

And as usual: feedback is welcome, especially the constructive kind :).

Comments

comments powered by Disqus

Archives