A general question about the use of classes in defining interfaces

[Ross Paterson]

On Wed, Oct 08, 2008 at 12:36:02PM +0200, S. Doaitse Swierstra wrote:
> Stimulated by remarks made during the discussion on the future of  
> Haskell at the last Haskell Symposium, I have started to convert my new 
> parsing library (constructed for the Lernet summerschool in Uruguay) into 
> Cabalised form. In this library I have amongst others the class:
>
> class  Applicative p where
>   (<*>)     ::   p (b -> a)  -> p b   ->   p a
>   (<|>)     ::   p a         -> p a   ->   p a
>   (<$>)     ::   (b -> a)    -> p b   ->   p a
>   pReturn   ::   a                    ->   p a
>   pFail     ::                             p a
>   f <$> p   =  pReturn f <*> p
>
> which extends/deviates from the standard class Applicative, since I  
> think these functions more or less belong together. I am happy to factor 
> out <|> into a separate class.

This corresponds to Alternative, a subclass of Applicative (except for <$>
being a function instead of a method).  They certainly belong together
for parsers, but there are applicative functors that don't have the
extra monoidal structure.

> The problem which arises now is when I want to use the class Applicative
> as it is now defined in Control.Applicative. Functions like <$>, <$, <*
> and many have standard implementations in terms of the basic function pure
> and <*>. Although this looks fine at first sight, this is not so fine if
> we want to give more specialised (optimised, checking) implementations,
> as I am doing in my library. An example of this is e.g. in many, where
> I want to check that the parameter parser does not recognise the empty
> sequence since thi is non-sense, etc. Of course we can describe <* by
>
> p <* q = pure const <*> p <*> q
>
> but this is also rather inefficient; why first building a result if you
> are going to throw it away anyway?

The current definition isn't quite as bad as that:

	p <* q = const <$> p <*> q
	f <$> a = fmap f a

but the general point stands.

> More in general I think it is preferred to place common patterns in  
> classes with a default implementation, so they can be redefined instead 
> of defining them at top level.
>
> 1) Does everyone agree with this observation, and if not what am I  
> missing?
> 2) Can we change the module Applicative to reflect my view? I think it  
> can be done without implying heavy changes to other modules.

This seems reasonable to me, as long as there aren't too many of them;
we do this for lots of other classes.  Would you like to list all the
functions that you would like to have as redefinable methods?