[Haskell-cafe] Re: Building "production stable" software in Haskell

[Benjamin Franksen]

Philippa Cowderoy wrote:
> On Mon, 17 Sep 2007, Adrian Hey wrote:
>> Ideally the way to deal with this is via standardised interfaces (using
>> type classes with Haskell), not standardised implementations. Even this
>> level of standardisation is not a trivial clear cut design exercise.
>> e.g we currently have at least two competing class libs, Edison and the
>> collections package. Which one should become standard?
> 
> They shouldn't, at least not now. Knock up something lightweight that'll 
> do for now for each of the modules that're going to be standard, worry 
> about overarching frameworks later. Realistically we need a standardised 
> name which we can expect to find an implementation under, with some 
> performance guarantees even if they're the worst possible ones we can 
> make.

I am using the collections package on a regular basis, and I am quite
satisfied. (I have no experience with Edison so I can't compare them.)

The main advantage of a framework such as the collections package offers is
that the code becomes a lot more flexible. First, it is easier to
experiment with different implementations. In one application I wrote, you
can switch from Data.Collections.StdMap (whose implementation is the
familiar Data.Map) to e.g. Data.Collections.AvlMap by changing
exactly /one/ line of code (of a total of about 1500 in 8 modules). No need
to change any of the import declarations, no change in type signatures,
nothing. Also, many functions, even whole classes, can be written more
polymorphically and thus easier to use in situations other than what they
were planned for; you can specify exactly the API that is needed and no
more which strengthens static typing.

The main disadvantage is that it can become quite hard to understand type
errors, which often don't give me any clue what the /cause/ of the problem
is. (This might be unavoidable, due to the level of polymorphism, I don't
know.)

In the end, I think something similar to the collections package should
become 'standard' in the sense of getting distributed with the main Haskell
implementations. This would encourage more people to try and use them, so
we'd gather more experience and would be able to eliminate shortcomings
sooner rather than later.

Cheers
Ben