Here’s a quick little note that has been bugging me for a while, and nobody wants to talk about it right now on IRC.
I think the By functions:
sortBy :: (a -> a -> Ordering) -> [a] -> [a] maximumBy :: (a -> a -> Ordering) -> [a] -> a groupBy :: (a -> a -> Bool) -> [a] -> [[a]] nubBy :: (a -> a -> Bool) -> [a] -> [a]
Etc. should be replaced by On functions:
sortOn :: (Ord b) => (a -> b) -> [a] -> [a] maximumOn :: (Ord b) => (a -> b) -> [a] -> a groupOn :: (Eq b) => (a -> b) -> [a] -> [[a]] nubOn :: (Eq b) => (a -> b) -> [a] -> [a]
My argument is: the functions provided to sortBy etc. have some preconditions. sortBy is not well-defined (or oughtn’t be) for functions which are not linear ordering functions; nubBy is shouldn’t be well-defined (to the dismay of some 31337s) for functions which do not encode an equivalence relation. But the folklore is that functions are typically “as total as possible”, so if it wants a function of some type, all I have to do is conjure a function of that type and my output will be something reasonable in terms of that function.
On the other hand, the folklore of typeclasses is that they typically come with laws. You need to prove — or at least think you know how to prove — some laws when you make a type an instance of a typeclass. The On functions use this obligation to encode their precondition. They are easier to use in a verified setting, too; there are a bunch of standard instances of Eq and Ord for which the laws are known to hold; map your data on to that in any way you like and the precondition is guaranteed.