Hi David, thank you very much for this proposal. I think having fail in Monad is just plain wrong, and I am therefore very happy to see it being moved out. I have some remarks, though:
A class of patterns that are conditionally failable are `newtype`s, and single constructor `data` types, which are unfailable by themselves, but may fail if matching on their fields is done with failable paterns.
The part about single-constructor data types is not true. A single-constructor data type has a value ⊥ that is different from applying the data constructor to ⊥’s. For example, ⊥ and (⊥, ⊥) are two different values. Matching ⊥ against the pattern (_, _) fails, matching (⊥, ⊥) against (_, _) succeeds. So single-constructor data types are not different from all other data types in this respect. The dividing line really runs between data types and newtypes. So only matches against patterns C p where C is a newtype constructor and p is unfailable should be considered unfailable.
- Applicative `do` notation is coming sooner or later, `fail` might be useful in this more general scenario. Due to the AMP, it is trivial to change the `MonadFail` superclass to `Applicative` later. (The name will be a bit misleading, but it's a very small price to pay.)
I think it would be very misleading having a MonadFail class that might have instances that are not monads, and that this is a price we should not pay. So we should not name the class MonadFail. Maybe, Fail would be a good name.
I think we should keep the `Monad` superclass for three main reasons:
- We don't want to see `(Monad m, MonadFail m) =>` all over the place.
But exactly this will happen if we change the superclass of (Monad)Fail from Monad to Applicative. So it might be better to impose a more light-weight constraint in the first place. Functor m might be a good choice. All the best, Wolfgang