I've been thinking a lot about the unsafePerformIO method of IO. I have two questions in particular that are on my mind. First, what would a hypothetical performIO method return? Second, why would it have to be an instance method of IO?

The answer to the first question almost suggests itself. While IO represents a deferred action that yields a result, with a possible alternative error state of Throwable, Try represents the result of performing an action which might be either a value or an error state of Throwable. So obviously a performIO on IO should return Try.

The second question has led me down a rabbit hole. This hypothetical performIO method is just a function from IO<A> to Try<A> with no opinion about A. It might as well be IO<?> to Try<?> except that it might be critical in a large number of use cases to unify the input and output carriers. Mathematically, this kind of function is a natural transformation, and shoki has a NaturalTransformation type that fits the mold, but it's kind of hard to use. I think I'd want it to have more convenient means of construction.

The only usage of the NaturalTransformation type is in the StreamT transformer, where it is the parameter to a mapStreamT method that turns a StreamT<M, A> into a StreamT<N, A> given a NaturalTransformation from the monad M to the monad N. The use is mystifying in its absence from other transformers. IterateT in lambda comes close with a mapIterateT that takes an Fn1 instead of a NaturalTransformation, but accomplishes the same goal of changing the argument monad of IterateT. Why don't other monad transformers offer this functionality? Why isn't StreamT's own mapStreamT method implemented as a NaturalTransformation? What other operations might be readily implemented in this form?

I'm starting to get motivated to fill this gap with code and see where it leads.