Haskell Typeclasses vs. C++ Classes
michaelburge.us
michaelburge.us
For onlookers, I'd note that Haskell typeclasses are something of a trap for beginners who think they are the way to accomplish object-like coding in Haskell.
Actually, records with functions (closures) inside is usually a better choice: simpler, more flexible, more straightforward.
Of course the resemblance between a vtable and a record of functions is very obvious!
The better choice would be to use Haskell as it was intended to be used.
If you need polymorphism you can mostly just reach for parametric polymorphism (generics) and closures. Existential types, in the form of closures, explicit existentially quantified types, or rank-2 universally quantified functions, give you basically the same expressive power as virtual functions—they’re all encodings of the same concept.
Typeclasses are more about the principle of “code to the interface, not the implementation”. For example, using a container through its Functor/Applicative/Monad/Monoid/Traversable/Foldable instances lets you be agnostic to its representation, swap it out for testing/performance, &c.
This. Group you code logically, expose only what's needed. This also gives the FPer a large part of the "encapsulation" and "group by responsibility" advantages that are often mentioned as pros of OO. Without going down the "implicit state sharing" path.
(0) Modules may expose n-ary operations on abstract types for n > 1.
(1) Modules may have multiple interrelated abstract type components, i.e., what Stepanov calls “multisorted algebras”.
Furthermore, if you are willing to use an impure language (like ML or Scala), modules can be just as stateful as objects if you deem it necessary.
As of C++17, concepts are not yet an actual language construct but are enforced by convention and normally only checked at template instantiation time (as opposed to definition time), although there are techniques that can be used to simulate more formal checking (archetypes, enableif, etc.).
Unfortunately concepts lite are not capable of type checking declarations, but it is still an improvement.
https://people.csail.mit.edu/dnj/teaching/6898/papers/wadler... (page 3)
Type-classes do allow interesting polymorphism tricks, that —as Wadler mentioned— allows Haskellers to use some of good parts of OO.
It is with cloud-native that it gets interesting. Cloud-native infrastructure (like k8s) usually replaces the traditional devops tools (like Puppet, Ansible and Salt). Also it brings the notion of "scrappable instances", "blue green deployments" and "stateless compute nodes". This is were it got really interesting, this is were the FP intuitions start carry over. :)
https://wycd.net/posts/2017-04-09-rpcs-for-moving-computatio...
s/not possible/possible, but highly discouraged/
Encapsulation — If you understand by OOP the encapsulation of data, then no, type classes have no such thing.
Inheritance — type classes don't have inheritance. What they have are constraints (e.g. you can implement typeclass X for type T if a typeclass implementation Y also exists for T). So you can say for example that MonadError can be implemented only for types already implementing Monad. That's not inheritance, but composition.
Most importantly perhaps and something that doesn't clearly follow from the above is that OOP as implemented in popular languages, including C++, gives you subtyping, whereas type classes do not.
This has important ramifications. For example downcasting isn't possible. And you no longer have the "liskov substitution principle". And subtyping is actually incompatible with Hindley-Milner type inference.
Also in actual usage the differences can be quite big — for example, if you view type classes as "restrictions" on what a type can do (like OOP interfaces are), you no longer need to add these restrictions on the types themselves, but on the operations, where you actually need them, while the data structure definition can remain free of such restrictions.
This is called "constrained parametric polymorphism" btw. Which can't be achieved in most OOP languages that I know of, except for Scala and that's because Scala has "implicit parameters" which are equivalent to Haskell's type classes.
At the end of the day, what a type class gives you is a way to transform a type name into a meaningful value. Imagine having a function like ...
f: T => A
data Object cls a where
Object :: cls a => a -> Object cls a
data Dict c where
Dict :: c => Dict
The more common terminology is ad-hoc polymorphism AFAIK.
data Nested a = Nest a (Nested [a]) | Done deriving Show
build :: Int -> a -> Nested a
build 0 a = Done
build i a = Nest a $ build (i- 1) [a]
main = do
i <- readLn
print (build i i)
Sorry that I mistyped the example at first. If you run this version:
> main
23
Nest 23 (Nest [23] (Nest [[23]] (Nest [[[23]]] (Nest [[[[23]]]] (Nest [[[[[23]]]]] (Nest [[[[[[23]]]]]] (Nest [[[[[[[23]]]]]]] (Nest [[[[[[[[23]]]]]]]] (Nest [[[[[[[[[23]]]]]]]]] (Nest [[[[[[[[[[23]]]]]]]]]] (Nest [[[[[[[[[[[23]]]]]]]]]]] (Nest [[[[[[[[[[[[23]]]]]]]]]]]] (Nest [[[[[[[[[[[[[23]]]]]]]]]]]]] (Nest [[[[[[[[[[[[[[23]]]]]]]]]]]]]] (Nest [[[[[[[[[[[[[[[23]]]]]]]]]]]]]]] (Nest [[[[[[[[[[[[[[[[23]]]]]]]]]]]]]]]] (Nest [[[[[[[[[[[[[[[[[23]]]]]]]]]]]]]]]]] (Nest [[[[[[[[[[[[[[[[[[23]]]]]]]]]]]]]]]]]] (Nest [[[[[[[[[[[[[[[[[[[23]]]]]]]]]]]]]]]]]]] (Nest [[[[[[[[[[[[[[[[[[[[23]]]]]]]]]]]]]]]]]]]] (Nest [[[[[[[[[[[[[[[[[[[[[23]]]]]]]]]]]]]]]]]]]]] (Nest [[[[[[[[[[[[[[[[[[[[[[23]]]]]]]]]]]]]]]]]]]]]] Done))))))))))))))))))))))
I'd say the difference is minor to non-existent.
There is no "isInstanceOf" check in Haskell.
C++ can do that with template specialization.
This might be nitpicking but vtables in C++ are created per type with instances having a pointer to it.
Although the language does not offer a builtin syntax for it, in C++ is also common to attach the 'vtable' to the 'handle' (a smart pointer or deep copying envelope), in the case of type erased wrappers (std::function, std::any, etc).
I enjoyed this post since it's kind of a mirror of this chapter, which was one of the hardest for us to figure out what to put in it.
1: https://doc.rust-lang.org/book/second-edition/ch17-00-oop.ht...
The trick here is defining "object-oriented". In modern object-oriented languages, I usually assume that involves encapsulation of state inside of boxes called objects. In that sense, type classes are quite different.
Methods on objects generally mutate or provide views into the contents of the box, while methods in type classes are just regular functions. Critically, there is no inner mutation or privileged view into data via type classes.
> They have things like properties and multiple inheritance
How do they have properties? Also, they don't have multiple inheritance - a type can only ever have one instance of any given typeclass.
> Even Wadler has a section leading with "Object-oriented programming."
I read Wadler's paragraph on object-oriented programming (bottom of page 3) more as a highlight of the fundamental _difference_ between type classes and object-oriented classes. The only similarity between typeclasses and classes are that they are the mechanisms by which runtime dispatch is achieved.
We should call programming with classes and inheritance. COP, as in Class Oriented Programming. COP is especially appropriate because if you don't follow the type rules, you go to compiler jail.
That is the Alan Kay definition. I'm not sure I'd claim that is the colloquial, universal understanding of OOP apart from the HN crowd, though. (ask a LOB programmer what OOP is and observe what they say).
I rather like the Alan Kay definition myself, but that is rather like having a Many Worlds understanding of quantum mechanics; Entirely consistent and arguably "correct", but not universally shared among practitioners.
We should refine our language (e.g. COP), but it would be revisionist to ignore what the 90's and 00's understanding of the definition would be.
I think it would be even more of a stretch to say that prototypal languages (javascript) are more OOP than non-prototypal languages, unless of course one is being a bit reductionist.
I'm not sure what your ordering relation is here, but if it has something to do with how "purely object-oriented" a language is, then, sure, JavaScript is more OOP than C++ and Java since classes are not first-class in the latter.
But Smalltalk, which is class-based, is even more OOP than JavaScript since in Smalltalk, constructor calls are simply regular method calls (unlike JS which has a special "new" keyword) and there are no primitive objects (unlike JS, which distinguishes between primitive types and their boxed forms, though it tries to obscure that fact).
> COP is especially appropriate because if you don't follow the type rules, you go to compiler jail.
Smalltalk, Ruby, and Python are all class-based OOP languages but do not have static type systems with lots of compiler errors.
This becomes even more apparent when using an actor-model language like Erlang. The types of "design patterns" you use focus much more around the communication and coordination you might find in real, physical systems. Consequently I find it easier to model "real" systems using actors than standard objects in a language like Java.
I also appreciate how the Pony language decomposes the notion of "objects" into actors that have their own process and capabilities (standard objects). It shows that their is a distinct role for both types of objects, and it helps when decomposing entities in a system.
If classes and inheritance aren't a form of OOP, then what are they?
Also, Smalltalk had classes and inheritance, whatever Alan Kay had to say on the matter.
Typeclasses have a lot more cognitive similarity to Go-style interfaces or Smalltalk protocols than they do to OOP-style classes, and that fact can be enormously confusing to newcomers (which, at times, seems to be the Haskell raison d'etre).
There are cases where that can't be done at all (polymorphic recursion, when the recursion can't be proven to terminate at compile time, I'm sure there are other) and runtime dispatching is required, but that's more of an exception. In C++ that's implemented with explicit, semi manually implemented, type erasure wrappers.
edit: as I said, I'm not an Haskell programmer; maybe in idiomatic Haskell the cases where runtime dispatch is required are much more common.
Comparatively, imagine you have a problem for which you need ad-hoc, but statically resolvable, overloads. In C++, you typically reach for templates. In Haskell, you instead typically reach again for typeclasses.
This is what I mean when I say neither comparison is really closer than the other. Haskell couldn't replace typeclasses with templates, just as it couldn't replace them with objects.
The whole world of GUI programming was based on OO.
The Xerox Star was introduced in 1981 https://en.wikipedia.org/wiki/Xerox_Star
Apple's Lisa was introduced in 1983. https://en.wikipedia.org/wiki/Apple_Lisa
With it, Lisa Toolkit and Clascal (Pascal with Classes) http://www.mirrorservice.org/sites/www.bitsavers.org/pdf/app...
The Mac in 1984 https://en.wikipedia.org/wiki/Macintosh_128K
MacApp in 1985 https://en.wikipedia.org/wiki/MacApp
The Tektronix 4404 running Smalltalk, also 1984 http://www.wirfs-brock.com/allen/files/tek/4404-Flyer.pdf
The Amiga and Atari ST in 1985.
Windows was at 2.0 in 1987.
Heck, the NeXT cube had been introduced by fall of '88.
By the time Haskell was first released, OOP was either already available in popular languages (Smalltalk, Object Pascal) or being introduced into existing languages (Objective C, C++, Lisp).
In hindsight it would appear there was no reason not to see that OOP was gaining in popularity, that the term "class" was already in use, and that overloading it would be confusing.
That said, in the end, all my original post was lamenting is that it's a shame that Haskell terminology went the way it did... it's just yet another aspect of the language that creates confusion among newcomers, and it already has enough of that already! :)
The haskell committee were concerned with conveying knowledge amongst functional programming researchers. The use of class in similar but unrelated fields doesn't mean everyone need use it.
Or you can think I'm just flying the mountain with ice cream windows if you like.
Lisp got away with an entirely different meaning for “car” for a pretty long time!
I'm saying in the broader context of programming languages, it's confusing. And that that confusion could have been easily foreseen and avoided... after all, Smalltalk had been using the term "class" for almost 20 years before Haskell appeared and introduced their terminology.
At least they're called "typeclasses" and not just "classes". But we're always going to have some overloading of these powerful short ancient words.
I did recently see a suggestion to rename the "typeclass" keyword to "algebra" which would be pretty cool!
I am not sure what other name you would suggest, anyway.
interface Eq a where
(==), (/=) :: a -> a -> Bool
x /= y = not (x == y)
implement (Eq a) => Eq (Tree a) where
Leaf a == Leaf b = a == b
(Branch l1 r1) == (Branch l2 r2) = (l1==l2) && (r1==r2)
_ == _ = False
They could have used a fashionable and somewhat overloaded term like "interface" or "protocol" or "trait" or "flavor", or they could have used an exact, albeit somewhat bland, name coming from mathematics. They did the latter.
In fact, they have faced a similar choice for many aspects of the language, and in most cases, mathematical purity got precedence over practicality. Had they chosen differently, the result wouldn't be Haskell.
Why not? They seem to be exactly interfaces to me.
> They could have used a fashionable and somewhat overloaded term like "interface" or "protocol" or "trait" or "flavor", or they could have used an exact, albeit somewhat bland, name coming from mathematics. They did the latter.
The terminology "class" in Haskell certainly has nothing to do with its mathematical meaning of "collection of sets defined by a predicate".
If we are talking about interface like in Java, for instance, then the important difference is that in Java, class specifies what interfaces it has, while in Haskell, any code can specify how a type is a member of a type class (actually there can be multiple different ways how the same type can be member of same type class). Another difference was, until recently, that the type classes can define "default" implementations of functions. And there are other differences related to "deriving" statement.
> The terminology "class" in Haskell certainly has nothing to do with its mathematical meaning of "collection of sets defined by a predicate"
I don't want to get into philosophy of math, I am not strong in it, but "class" in mathematics is certainly not limited to collections of sets. For example, there is a class of all classes, but not set of all sets.
The predicate that defines the type class in Haskell is defined by the collection of all the relevant "instance" statements.
Actually, now that I think about it, as I note above, there are multiple ways that one type can be made a member of one type class, so it's certainly not a membership in the set-theoretic sense.
Let's say you have some function called GetStringableThing() that returns returns, as its name suggests, some object that implements the Stringable interface -- where Stringable is just an interface that consists of a ToString() method. As the caller of GetStringableThing(), you have no say in precisely which type of Stringable thing you'll get -- that's up to the callee. So there exists some type that implements Stringable, which GetStringableThing() would be more than happy to give you. This is in contrast with universal quantification; if the return type of GetStringableThing() was universally quantified, that would mean that you as the caller get to dictate the return type (so long as it implements Stringable), and its the callee's responsibility to be able to return any type the caller demands.
Typeclasses (as in Haskell) have the benefit of allowing you, as the caller of a function, to dictate the return type of that function -- in fact, universal quantification is the default, and you have to go out of your way (e.g. with GADTs) to represent existential types.
This now leads us to a common pitfall for people coming to Haskell from OOP languages, where their intuition that "typeclass == interface" fails them:
-- The intent here is the same as earlier with "Stringable",
-- though in Haskell that would be "Show".
--
-- A beginner's intuition:
-- "Surely this reads 'somethingShowable is some type implementing the Show interface', right?"
somethingShowable :: (Show a) => a
somethingShowable = (42 :: Int)
This bit of OOP pseudo-code captures what they we're going for:
func getStringableThing() Stringable {
return 42
}
Typeclass instances are resolved at compile time (except where you've gone out of your way to existentially type something), and can be specialized and inlined accordingly. Interfaces (a la OOP) are implemented via some sort of vtable mechanism, where dispatch is deferred until runtime (though a JIT can potentially do some inlining).
Interfaces (a la OOP) and typeclasses are implemented (and behave) quite differently, so I would argue that if Haskell were to replace its use of "class" (and "typeclass") with "interface" it would only confuse things further (as evidenced by the constant stream of beginners asking for help where they've made this exact conflation; for the record, I was one of these beginners 3 years ago).
The "they shouldn't call these (type)classes!" complaint is quite tiring: no one ever has a reasonable alternative to offer. At least (type)class makes some sense when you realize that "class" is meant in the mathematical sense of the word "class".
It's still confusing. And programming languages exist to first and foremost aid programmers, not to adhere to some abstract expression of mathematical perfection.
At the time Haskell hit the scene, the term Interface was free and could've been taken. But that's just one thought, I'm not a language designer...
(given) g :: Int -> String
function a =
let
l = (f a)
in
g l
"function :: (a which has a method f :: a -> Int) -> String"
Now you're not saying that a has to belong to a typeclass, you are saying that a has a duck-type of any value with a method f :: a -> Int.
If I understood correctly, the OP asked for a “duck type of all values on which a function `f` can be applied”. Alas, OCaml-like object types are glorified function types, not duck types:
(0) The type of an object is the type of its methods.
(1) The type of an object is not the type of its internal state.
Say you have a particular data type like `float list`, and you want to “bless it with a method `f`”, so that you can pass a `float list` to any context that expects an object with a method `f`. This is everyday stuff in duck-typed languages.
Turns out, you can't. The best thing you can do is explicitly construct an object with a method `f`, and whose state happens to be a `float list`. Alas, because the state isn't a part of the object's type, there is no way a user of the object can recover the original `float list`. In fact, no user of the object has any way to tell whether the object's internal state is a `float list` or something else altogether.
I think that's the other way around. It appears to be what they asked for, even if it may not be what they want.
They pretty clearly said they wanted to be able to write a type that says "Object that has method f :: a -> Int".
What you're talking about doesn't seem to be duck typing (though it might work nicely along side duck typing), it's being able to add methods to objects at run time (unless I'm misunderstanding).
This is pretty succinct, but I'd have to say that it is not obvious what the syntax means:
val y : < get_x : int; set_x : int -> unit > = <obj>
So this:
val y : < get_x : int; set_x : int -> unit >