Two ways of currying in Scala; what's the use-case for each?

Solution 1:

Multiple Parameter List Methods

For Type Inference

Methods with multiple parameter sections can be used to assist local type inference, by using parameters in the first section to infer type arguments that will provide an expected type for an argument in the subsequent section. foldLeft in the standard library is the canonical example of this.

def foldLeft[B](z: B)(op: (B, A) => B): B

List("").foldLeft(0)(_ + _.length)

If this were this written as:

def foldLeft[B](z: B, op: (B, A) => B): B

One would have to provide more explicit types:

List("").foldLeft(0, (b: Int, a: String) => a + b.length)
List("").foldLeft[Int](0, _ + _.length)

For fluent API

Another use for multiple parameter section methods is to create an API that looks like a language construct. The caller can use braces instead of parentheses.

def loop[A](n: Int)(body: => A): Unit = (0 until n) foreach (n => body)

loop(2) {
   println("hello!")
}

Application of N argument lists to method with M parameter sections, where N < M, can be converted to a function explicitly with a _, or implicitly, with an expected type of FunctionN[..]. This is a safety feature, see the change notes for Scala 2.0, in the Scala References, for an background.

Curried Functions

Curried functions (or simply, functions that return functions) more easily be applied to N argument lists.

val f = (a: Int) => (b: Int) => (c: Int) => a + b + c
val g = f(1)(2)

This minor convenience is sometimes worthwhile. Note that functions can't be type parametric though, so in some cases a method is required.

Your second example is a hybrid: a one parameter section method that returns a function.

Multi Stage Computation

Where else are curried functions useful? Here's a pattern that comes up all the time:

def v(t: Double, k: Double): Double = {
   // expensive computation based only on t
   val ft = f(t)

   g(ft, k)
}

v(1, 1); v(1, 2);

How can we share the result f(t)? A common solution is to provide a vectorized version of v:

def v(t: Double, ks: Seq[Double]: Seq[Double] = {
   val ft = f(t)
   ks map {k => g(ft, k)}
}

Ugly! We've entangled unrelated concerns -- calculating g(f(t), k) and mapping over a sequence of ks.

val v = { (t: Double) =>
   val ft = f(t)
   (k: Double) => g(ft, k)       
}
val t = 1
val ks = Seq(1, 2)
val vs = ks map (v(t))

We could also use a method that returns a function. In this case its a bit more readable:

def v(t:Double): Double => Double = {
   val ft = f(t)
   (k: Double) => g(ft, k)       
}

But if we try to do the same with a method with multiple parameter sections, we get stuck:

def v(t: Double)(k: Double): Double = {
                ^
                `-- Can't insert computation here!
}

Solution 2:

You can curry only functions, not methods. add is a method, so you need the _ to force its conversion to a function. add2 returns a function, so the _ is not only unnecessary but makes no sense here.

Considering how different methods and functions are (e.g. from the perspective of the JVM), Scala does a pretty good job blurring the line between them and doing "The Right Thing" in most cases, but there is a difference, and sometimes you just need to know about it.

Solution 3:

I think it helps to grasp the differences if I add that with def add(a: Int)(b: Int): Int you pretty much just define a method with two parameters, only those two parameters are grouped into two parameter lists (see the consequences of that in other comments). In fact, that method is just int add(int a, int a) as far as Java (not Scala!) is concerned. When you write add(5)_, that's just a function literal, a shorter form of { b: Int => add(1)(b) }. On the other hand, with add2(a: Int) = { b: Int => a + b } you define a method that has only one parameter, and for Java it will be scala.Function add2(int a). When you write add2(1) in Scala it's just a plain method call (as opposed to a function literal).

Also note that add has (potentially) less overhead than add2 has if you immediately provide all parameters. Like add(5)(6) just translates to add(5, 6) on the JVM level, no Function object is created. On the other hand, add2(5)(6) will first create a Function object that encloses 5, and then call apply(6) on that.