Functional debugging in Scala

October 31, 2013

Printing to standard output might be one of the earliest debugging techniques we discover. We write a program, sprinkle some println statements here and there, and look for where the program fails or what its runtime values are.

I still occasionally find myself reaching for this technique, sometimes in the form of a logging framework, to get basic runtime inspection of a program, but I find it hard to correlate log messages with program state, and even harder to consistently reproduce across development, testing, and production environments.

Fortunately, the writer monad offers a way out of this bind - we can have our logging and maintain referential transparency too!

Let's start with a side-effectey example. Consider the following arithmetic functions:

val add: Int => Int => Int =
  x => y => x + y

val mult: Int => Int => Int =
  x => y => x * y

val div: Int => Int => Option[Double] =
  x => y => if (y == 0) None else Some(x.toDouble / y)

val parse: String => Option[Int] =
  x => try { Some(x.toInt) } catch { case t: Throwable => None }

The add and mult functions both take two integers and produce another integer. The div and parse functions, which can fail (since they're only partially defined over their domains), both produce an Option of a number.

We can string these functions together, taking advantage of Option's map and flatMap functions, to perform a compound operation:

for {
  x1 <- parse("42")
  x2  = mult(x1)(2)
  x3  = add(x2)(42)
  x4 <- div(x3)(3)
} yield x4  // returns Some(42.0)

We can also interleave some debugging statements using println:

for {
  x1 <- parse("42")
   _  = println("x1: " + x1)  // prints "x1: 42" to stdout
  x2  = mult(x1)(2)
   _  = println("x2: " + x2)  // prints "x2: 84" to stdout
  x3  = add(x2)(42)
   _  = println("x3: " + x3)  // prints "x3: 126" to stdout
  x4 <- div(x3)(3)
   _  = println("x4: " + x4)  // prints "x4: 42.0" to stdout
} yield x4                    // returns Some(42.0)

But we want to avoid both the side-effect of printing, and the dissociation of the result data from the log messages.

Let's build a special logging data structure that, when composed with Option instances, allows us to keep the same shape of our for comprehension, but return the log (along with the final result) as a sequence of log messages:

class LogOption[A](val run: Option[(A, Seq[String])]) {
  def map[B](f: A => B): LogOption[B] =
    new LogOption(run map { x => (f(x._1), x._2) })
  def flatMap[B](f: A => LogOption[B]): LogOption[B] =
    new LogOption(run flatMap { case (a, l) =>
                   f(a).run map { case (b, l2) => (b, l ++ l2) } })
}

implicit def logOption[A](x: Option[A]): LogOption[A] =
  new LogOption(x.map(a => (a, Nil)))

def log(x: String): LogOption[Unit] =
  new LogOption(Some(((), Seq(x))))

The LogOption class wraps a function that returns an optional pair of a result plus a log. Such a result could be as simple as:

Some((42.0, "returning 42.0" +: log))

The LogOption class specifies how (via map) to apply a function to its result type, as well as how (via flatMap) to compose itself with a function thet returns another LogOption.

We can use it with only minor modification to the code above:

val x = for {
  x1 <- logOption(parse("42"))  // lift Option[Int] to LogOption[Int]
  _  <- log("x1: " + x1)
  x2  = mult(x1)(2)
  _  <- log("x2: " + x2)
  x3  = add(x2)(42)
  _  <- log("x3: " + x3)
  x4 <- logOption(div(x3)(3))   // lift Option[Double] to LogOption[Double]
  _  <- log("x4: " + x4)
} yield x4
x.run  // returns Some((42.0,List(x1: 42, x2: 84, x3: 126, x4: 42.0)))

Now we have a way to compose functions that return raw integers and Options of integers, while building up a queue of log messages. Nothing is written to standard output, no external state is altered, and in fact the code isn't even executed until it is initiated with an empty log via x.run(Nil).

It turns out that LogOption is a specialization of the writer monad transformer:

case class WriterT[F[_],W,A](run: F[(A,W)])
type Writer[W,A] = WriterT[ID,W,A]

type WriterTM[F[_],W,A] = Monad[({type λ[α] = WriterT[F,W,α]})#λ,A]

implicit def writerT[F[_],W,A](x: WriterT[F,W,A])
    (implicit liftM: LiftM[F], liftS: W => Semigroup[W]): WriterTM[F,W,A] =
  new WriterTM[F,W,A] {
    def run: F[(A,W)] = x.run
    def map[B](f: A => B): WriterT[F,W,B] =
      new WriterT[F,W,B](liftM(x.run) map { x => (f(x._1), x._2) })
    def flatMap[B](f: A => WriterT[F,W,B]): WriterT[F,W,B] =
      new WriterT[F,W,B](
        liftM(x.run) flatMap { case (a,w1) =>
          liftM(f(a).run) map { case (b,w2) => (b, w1 * w2) } })
  }

type LogT[F[_],A] = WriterT[F,Seq[String],A]

def logT[F[_],A](x: F[A])(implicit lift: F[A] => Monad[F,A]): LogT[F,A] =
  new LogT(x.map(a => (a, Nil)))

def log(x: String)(implicit lift: Seq[String] => Semigroup[Seq[String]]): LogT[Option,Unit] =
  new LogT(Some(((), Seq(x))))