[Haskell-cafe] IO (Either a Error) question

[Ben Millwood]

On Sat, May 8, 2010 at 3:26 AM, John Meacham <john at repetae.net> wrote:
>
> What counts as unfailable?
>
> (x,y) probably,  but what about
>
> data Foo = Foo x y
>
> If we don't allow it, we add 'magic' to tuples, which is a bad thing, if
> we do allow it, there are some odd consequences.
>
> adding another constructor to Foo will suddenly change the type of do
> notations involving it non locally. said constructor may not even be
> exported from the module defining Foo, its existence being an
> implementation detail.
>
> All in all, it is very hacky one way or another. Much more so than
> having 'fail' in Monad.
>

This is an interesting point, but I still disagree. A data type having
constructors added or changed *is* going to break code in clients
using it, or at least make GHC spit out a bunch of non-exhaustive
warnings. It's then a good idea, I think, that people are forced to
re-examine their use sites which don't obviously handle the new
failing case. Presumably if they were really really sure then just a
few well-placed ~s would make the problem go away.
(i.e. to answer your question, pattern matching against any
single-constructor data type should be unfailable in my opinion).

On Sat, May 8, 2010 at 7:16 AM, Ivan Lazar Miljenovic
<ivan.miljenovic at gmail.com> wrote:
> As I said in another email: does not the "x <- Nothing" itself call fail
> as it expects x to be an actual value wrapped in Just?

No, the propagation of Nothings is done solely by the definition of
>>= for Monad, and doesn't need fail at all.