Practical Future[Option[A]] in Scala

Posted on March 7, 2014

Motivation

In real world concurrent code you often come across the Future[Option[A]] type(where A is usually some concrete type). And then you need to compose these things. This is not straightforward in Scala.

If you’ve done some Haskell just import scalaz._ and you can skip the rest of this article. Scalaz library defines a monad typeclass(and many others) that formally specifies what it means to be a monad(not “has a flatMap-ish thingy”). Then it’s easy to build abstractions upon this.

But what if you don’t want to add another dependency and just want to make this tiny bit more practical? I can be done and is not that complicated. You will also learn something and maybe even become motivated to bite the bullet and start using Scalaz.

Meaning

The Future[Option[A]] type combines the notion of concurrency(from Future) and the notion of failure(from Option) giving you a concurrent computation that might fail. What we want is a monad instance for this. With only standard library code you probably would use Future#flatMap in combination with Option#map and Option#getOrElse(or just Option#fold). This gets messy and unreadable quite quickly due to loads of boilerplate. Let’s fix this!

Newtyping

A monad in scala means a flatMap method, so it’s safe to assume we need to define a flatMap with this special semantics somewhere. You might want to define an implicit class that has the new method but it wouldn’t work as Future already has a flatMap method. I took a page from Haskell’s book. When you want to override semantics there you wrap up the type into a newtype. This is just compile-time type information that is completely free at runtime. Luckily Scala has this in form of AnyVal.

import concurrent.Future

case class FutureO[+A](future: Future[Option[A]]) extends AnyVal

I’ve also made it covariant for ease of use.

The Monad

It’s actually quite easy to implement

import concurrent.{Future, ExecutionContext}

case class FutureO[+A](future: Future[Option[A]]) extends AnyVal {
    def flatMap[B](f: A => FutureO[B])
                  (implicit ec: ExecutionContext): FutureO[B] = {
        FutureO {
                future.flatMap { optA => 
                optA.map { a =>
                    f(a).future
                } getOrElse Future.successful(None)
            }
        }
    }

    def map[B](f: A => B)
              (implicit ec: ExecutionContext): FutureO[B] = {
        FutureO(future.map(_ map f))
    }
}

You need to pull in an execution context and then do the usual boilerplate thing ending up with a wrap again. I’ve also added map method which is trivial but we need it because in Scala for comprehensions are desugared into flatMap and map to avoid using point(abstract constructor) since it’s harder to express OO-style..

Usage

What good is this FutureO? Let’s do a contrived example. We need a function that might fail - divideEven that only divides even numbers. And we need concurrency - we’ll divide two numbers concurrently.

def divideEven(n: Int): Option[Int] = 
    if (n % 2 == 0) Some(n/2) else None

//first spawn both computations
val f1 = Future(divideEven(14))
val f2 = Future(divideEven(16))

//and combine them
val fc = for {
    a <- FutureO(f1)
    b <- FutureO(f2)
} yield a + b

//prints out Success(Some(15))
fc.future onComplete println

It works. How would this look without FutureO?

val fc = for {
    oa <- f1
    ob <- f2
} yield for {
    a <- oa
    b <- ob
} yield a + b

fc onComplete println

Two layers of for comprehensions. It gets even hairier if you have data dependencies between your futures. Consider divideEven again but this time we want to divide a number twice in a row. And we’ll be keeping the futures around just to prove a point. Let’s imagine that divideEven does some blocking IO and we want to push it into another tread-pool.

def divideTwiceF(n: Int): Future[Option[Int]] = {
    val fo = for {
        n1 <- FutureO(Future(divideEven(n)))
        n2 <- FutureO(Future(divideEven(n1)))
    } yield n2
    fo.future
}

And it works as expected. The FutureO part there is just to alter the monadic semantics. As an exercise try to rewrite this without FutureO and squirm in disgust.

Usability

Inside for comprehension(or manual flatMaps) you can still construct failed futures, throw or return None(in a future). However putting in a default value(getOrElse) is a bit trickier and going back to regular Future inside same comprehension is impossible. But you can fix this. You can define methods like orElse on the FutureO. You can also overload the flatMap to enable interop with regular futures. However this screws up type inference and I would advise against it as it could introduce some nasty bugs.

Try to implement combinators you need and leave some comments. Especially if you come across something nice or find a case where FutureO is more awkward to use than regular futures.

Theory

What we defined is actually a specialized monad transformer. Monads(such as Future and Option) have this nasty property that they don’t compose. You cannot write a function that takes a two monads and outputs a composed one.

However you can write such a function if you fix one of the monads and only take one as a parameter. This is called a monad transformer. In this case I fixed Option. Take a closer look, we are only using flatMap and a constructor(point) from Future. This means we could abstract over the whole monad class. And this is what Scalaz does with OptionT. But to do this it needs to define a monad typeclass and instances for each and every monad they find. What I did is fix the other monad too and this produces a concrete instance you can use without any typeclasses. There is a downside of course. This is a one-off hack. If you want other transformers you’ll have to write them as well. At that point I think would be a good idea to start using Scalaz.

Facebook Twitter Google Digg Reddit LinkedIn StumbleUpon Email
comments powered by Disqus