Hacker News Reader

Currying is fun and I miss it sometimes as well, but as near as I can tell, it is essentially immiscible with traditional languages. I've seen so, so many attempts, and they are so, so bad.

Surprisingly few attempts even seem to get over the bar of understanding what currying is, and why it isn't just partial application. Now, I am not too concerned with which is "better" or anything, nor do I much care for pedantry for the sake of pedantry, but in this case it's a matter of type signatures; curry(some_function) and partial_apply(some_function, param1, param2) are fundamentally different type signatures and for that reason, not because of abstract verbal purity or something, aren't the same thing.

I mean, right out of the gate you have the problem that in most conventional languages you're looking at your first example being map(reverse)(["hello", "goodbye"]).

Whether Ruby syntax could be bent into making it work without parens (or Perl perhaps) I don't know but with all the other things going on in those syntaxes I'm sure it would be quirky as can be, possibly quirky beyond practical usefulness. And in dynamic scripting languages you'd be looking at a lot of actual function calls, so performance is going to be trashed if you use this pervasively.

I agree that the Haskell approach is not easily portable to other languages. At least according to the article, the Ruby community is struggling to find a solution. I don't think parentheses (or their absence) is a decided factor: for example, I don't find the hypothetical Ruby syntax ["hello","goodbye"].map(&reverse) to be offensive or wildly inconsistent. Note that the map function itself doesn't need partial application, only its function parameter.

For the same reason, I disagree that performance in dynamic scripting languages would suffer unduly.

> I don't find the hypothetical Ruby syntax ["hello","goodbye"].map(&reverse) to be offensive or wildly inconsistent.

This syntax wouldn't work in Ruby, because bare `reverse` is already a method call, not a reference to a method by name. Allowing method calls without parentheses is crucial to Ruby's design, where all objects are fully opaque, and every `obj.attr` looking like a getter, is just a call to an instance method `attr`. (This is, as far as I understand, the opposite to similar languages like Python/JS, where the object is a dictionary of attributes, and `obj.method` reference the attribute of type "method", while `obj.method()` is an invokation.)

This design became a huge drawback in the age of functional(ish) programming, because the simplest way for refer to a method in Ruby is `method(:its_name)` (which is also inefficient, because it creates wrapping object of class Method on the fly, there is no pre-existing first-class object), and any attempt of passing/combining methods would be cumbersome due to it.

FWIW, you can do that in Ruby:

  (JSON.method(:parse) >> method(:puts)).call("[1, 2, 3]")

...which is semantically cool, but looks ridiculous.

Another important trait of Ruby's design is that most of the important methods belong to their first argument, so it is not `reverse(string)` but `string.reverse`, so even to obtain a reference to a method, you need to have an object it belongs to! So you can't do that:

   method(:>).call(a, b)

...because all operators are called on their first operands, so you need this this:

   a.method(:>).call(b)

...which will require to refer to at least a first argument of the operator, so you can't produce "argument-less comparison operation to call later".

This is the sort of thing I meant by "essentially immiscible". At a 30,000 foot view, yeah, it looks like lots of languages could "just" add currying.

But when you get down into the weeds of what it means, it turns out to be something you can't "just" drop in. You really need to build the syntax for it from day one.

> you can't produce "argument-less comparison operation to call later".

You absolutely can, if you know the specific one (that is, type) you need.

E.g.,

  gt = Integer.instance_method(:>).method(:bind_call)
  # gt is a two arg callable functionally equivalent to ->(a,b) { a < b } for Integer a, Any b

Of course, doing this instead of just doing:

  gt = -> { _1 > _2 }

loses Ruby duck-typing and dynamism with regard to the first parameter; which may be acceptable in some cases, but not others. Can you do the general form without an explicit lambda? Probably there is a way, but its not easy.

  Surprisingly few attempts even seem to get over the bar of understanding what currying is, and why it isn't just partial application.

Huh, what is the difference between currying and partial application? I always thought the terms were interchangeable.

> Huh, what is the difference between currying and partial application?

Partial application (of one arg) takes a function that is of type:

   (a, b, c) -> d

And makes it:

  a -> (b, c) -> d

Currying makes it:

  a -> b -> c -> d

Ruby’s Proc#curry does magic that mostly lets you pretend its either or something in between.

> In Haskell, by way of comparison, any one-parameter function can be used as the argument to map

Ruby doesn’t have functions, but any one parameter block (to which any one parameter callable, and some other things, can be converted with an &) can be passed as the block parameter to map. And, yes, you can use function composition and similar operations to build up more complex callables from existing ones without an explicit new block or lambda.

Your examples mostly have pretty close parallels in ruby:

  irb(main):044:0> ["hello", "goodbye"].map &:reverse
  => ["olleh", "eybdoog"]

is the more compact form of the form with the explicit lambda:

  irb(main):045:0> ["hello", "goodbye"].map {|s| s.reverse}
  => ["olleh", "eybdoog"]

composition works similarly, though there is no compact notation to convert an operator and its right operand to a callable, only the left:

  irb(main):048:0> ["a","bb","ccc","dddd"].filter &(2.method(:<) << :length.to_proc)
  => ["ccc", "dddd"]

But, you do have to cheat a little to do:

  irb(main):049:0> [[1,2],[3,4]].map &:*.to_proc.uncurry
  => [2, 12]

in place of:

  irb(main):049:0> [[1,2],[3,4]].map {|x, y| x * y}
  => [2, 12]

Because while ruby has Proc#curry, it lacks Proc#uncurry. But its simple to define:

  class Proc
    def uncurry
      ->(args) { self.call *args }
    end
  end

> ghci> filter (\x -> (length x) > 2) ["a","bb","ccc","dddd"] ["ccc","dddd"]... I find the Haskell syntax concise and clear...

> ["a","bb","ccc","dddd"].filter &(2.method(:<) << :length.to_proc)

I can't imagine a stronger case for GP's thesis. The point-free ruby rewrite is so baroque and needlessly circuitous... the extra variable is a far lesser evil, and more concise:

    ["a","bb","ccc","dddd"].filter { |x| x.length > 2 }

my instinct is

    ["a","bb","ccc","dddd"].grep /.../

but then, I'm a recovering perl hacker

Love it!

Yeah, mirroring the Haskell code, while possible in Ruby, is definitely not the best Ruby-aware approach, either for concision or clarity. The composability is available, but the problems where it is the best solution are different.

Even without patching the standard library:

    irb> [[1,2],[3,4]].each_with_object(:*).map(&:reduce)
    => [2, 12]

Given the mechanics of this I wouldn't claim it meets a strict definition of point-free, but that's an iterated application of operator-like methods omitting any explicit named/positional argument references.

I made a small nod towards Haskell's way of doing things in the last paragraph of "How others do it" section:

> We might also start look into the concept of the “tacit programming”, which, from some point of view, is also about “not repeating the arguments,” but this would be a much longer post—while it is obscenely long already.

Note though that my analysis was mostly about "how you would do that in an established language," not "how would I design a language from scratch." And Ruby is very different from Haskell in its syntax, semantics, and typical intuitions of the language writer and reader.

So designing a solution for clearer representation of "default parameters" is quite different in a language which is built around an `object.method(argument) { block }` as its atomic phrase, and one built around `function(arguments)`. (This can turn into a discussion of which phrase structure is more elegant in general... But I believe it was already done quite a few times in a history of programming languages evolution!)

I don't know how people feel about point-free these days. It was cool at first, then hated quite a lot due to issues I forgot (maybe a case or write-only code practice), but I find there's a high technical value and poetry to craft terms that plug together well with near no linking/binding declaration.

"Just passing a function" is incompatible with Ruby's current syntax where a bare method name calls the method unless you then introduce a way for method definitions to defer evaluation of its arguments, but then you open a can of worms given that Ruby allows aliasing and replacing arbitrary methods of already defined classes. I suspect it'd be a nightmare of inconsistencies.

It’s a bit pedantic to mention, but there is #method(:method_symbol), and obviously &:method_symbol for some cases. Both of which give you a proc you could call later with arguments provided to it. It’s not “the function” since, it’s really an anonymous function, bound to something, that will send :method_symbol to that thing. Gets around the bare-method-runs-method and metaprogramming issues with current Ruby syntax.

The entire point was that those constructs that do not just pass the bare method or something like a Proc that allows you to call the method, are needed exactly because the syntax will not allow for just mentioning a bare method name without calling it.

If you're fine with the extra syntax, yes, there are many alternatives, but they don't address the point.

Ruby does allow you to pass a one-parameter function as the argument to `map`:

  irb(main):001:0> ["foo", "bar"].map(&:reverse)
  => ["oof", "rab"]

Not precisely the same as what you've outlined, of course, but perhaps useful nonetheless.

Fun fact, the arity of functions generated by Symbol#to_proc is not 1 but -2 (that is, additional arguments are passed through as well), so we can write stuff like

    module MissionControl
      #...
    
      [Proc, Method, Binding, UnboundMethod].each_with_object(self, &:include)
    end

if one were so inclined.

Huh - truly did not know about _negative_ arity in Ruby, and what that actually means. I learned something today, thank you!

Just my opinion here...

The first example with map reverse is very readable and an improvement over lambda syntax. The chaining example (filter on length) is horrible for readability and I greatly prefer the lambda syntax.