Why Scala's Traversable Is Bad Design

| 4 minutes | Comments

Given there’s a Scala collection redesign discussion going on, it’s a good time to talk about one of my personal pet peeves: the existence of Traversable in the standard library, along with its variants like TraversableLike and TraversableOnce. Apparently this interface is missing in the new design and that’s awesome.

It’s easy to make API mistakes, we all do it and it’s important to learn from past mistakes, this document serving as a lesson for why Traversable is a bad idea.

Claims:

  1. Traversable has implicit behavior assumptions that are not visible in its exposed signature, the API being error prone
  2. Iterating over a Traversable has worse performance than Iterator
  3. There exists no Traversable data type that doesn’t admit an efficient Iterator implementation, thus Traversable being completely redundant

As a reminder and you can also read the docs, the Traversable is a trait like the following:

trait Traversable[+A] {
  def foreach(f: A => Unit): Unit
}

The standard library also has the venerable Iterable / Iterator:

trait Iterable[+A] {
  def iterator(): Iterator[A]
}

trait Iterator[+A] {
  def hasNext: Boolean
  def next(): A
}

Can you spot the similarities?

You should, because these 2 interfaces are supposed to be duals. So if you think of Traversable as being defined by that foreach function, then Iterable is that function with its arrows reversed:

type Traversable[A] = (A => ()) => ()

type Iterable[A] = () => (() => A)

Now this is interesting. For one Traversable is a sort of inversion of control technique, so instead of having a cursor that you have to manually advance, you now register a callback to a function and that callback gets called for you on each element. This actually frees us from certain Iterator constraints. For example with a push-based API we should no longer care when those function calls happen.

But you should already spot problems with the above definition. Our Iterable function signature isn’t complete, this one is:

type Iterable[+A] = () => Iterator[A]

type Iterator[+A] = () => Try[Option[A]]

Or in other words any Iterator can:

  1. give us the next element,
  2. or signal completion or failure

This means that the actual dual of Iterator is:

type Observer[A] = Try[Option[A]] => Unit

Or for those OOP-oriented among us, I give you the Observer as championed by Rx.NET, as the true dual of Iterator:

trait Observer[-A] {
  def onNext(a: A): Unit
  def onComplete(): Unit
  def onError(ex: Throwable): Unit
}

(Hello Monix :-))

This matters because Traversable has no way to signal completion or failure, unless you get a guarantee that all the processing happens synchronously, everything being over after the invocation of its foreach.

As an abstraction, this makes it useless when compared with Iterable and Iterator. If you introduce the synchronous execution constraint, there exists no data type that can implement Traversable and that doesn’t admit an Iterator implementation. None.

Even more problematic in my opinion is that this restriction isn’t visible in its API, unless your eyes are trained for it. With Iterator.next() whether you want it or not, you have to process things synchronously, because the signature says so.

Also problematic is TraversableOnce, which is supposed to be a traversable that can only be traversed once, like its name says. We’ve got this:

trait TraversableOnce[+A] {
  def foreach(f: A => Unit): Unit
}

trait Traversable[+A] extends TraversableOnce[A]

Besides the name and the inheritance relationship, there is no difference. This is another problem. Even if the API is effectful/impure, we should still be able to use types serving as documentation. Contrast with the Iterable / Iterator separation. Iterating over an Iterator is known to consume it and you can see this in its API. And the generator/factory part is in Iterable, which is good separation of concerns.

Traversable also has worse performance than Iterator. The ugly truth is that the JVM hasn’t been doing a good job at inlining that function reference you pass to foreach. This is called the inlining problem, which happens for megamorphic function calls in hot inner loops.

So there you have it and I hope that along with the redesign we’ll get rid of Traversable.

My opinions have been highly influenced by the work of Erik Meijer, if you want to learn more checkout this presentation: Contravariance is the Dual of Covariance Implies Iterable is the Dual of Observable.

| Written by
Tags: Scala