Programs as Values, Part V: Outputs

We want to write an algebra that reads from stdin, and use it to write the following program:

read a line from stdin, compute its length, and return true if the length is greater than 10, or false otherwise.

A starting point could be:

/*
 * carrier:
 *   In
 * introduction forms:
 *   readLine: In
 */
sealed trait In {
  ...
object In {
  val readLine: In
  ...

but obviously that’s not enough to write our program: we have encoded the action of reading a line from stdin, but we still need to do some extra transformations on the line we’ve read.

So, your first instinct might be to add an elimination form:

/*
 * carrier:
 *   In
 * introduction forms:
 *   readLine: In
 * elimination forms:
 *   nope: In => String
 */
sealed trait In {
  def nope: String
  ...
object In {
  val readLine: In
  ...

in order to write:

val out: Boolean = readLine.run.length > 10

but this is not a fruitful direction: we are basically saying that the only way to change the output of a program written in the In algebra is to eliminate the algebra entirely.

Algebras are our unit of composition, therefore with the elimination approach any program that needs to change its output can no longer be composed, which is a really strong limitation: for example in Part III we saw that for IO elimination happens when the JVM calls main, it would be really weird if we couldn’t encode something as simple as String.length until then.

Instead, we want to have the ability to transform outputs without leaving our algebra, and therefore we have to enrich it with a transformOutput combinator. Recall that the general shape of a combinator is:

transformOutput: (In, ...) => In

and we need to fill the ... with something that can encode the idea of transforming one thing into another, which we already have a well known concept for: functions. So, transformOutput needs to take a function, but we have a problem: what type should this function be, in order to fit the possible transformations we want to encode such as _.length or _ > 10 ?

Of course, an Any => Any fits anything:

/*
 * carrier:
 *   In
 * introduction forms:
 *   readLine: In
 * combinators:
 *   transformOutput: (In, Any => Any) => String
 */
sealed trait In {
  def transformOutput(transform: Any => Any): In
  ...
object In {
  val readLine: In
  ...

but this is also not an acceptable solution: our algebra has gained power, but we have lost type safety altogether.

As it turns out, the issue is with the carrier, specifically that these two programs have the same type:

val readsLine: In
val computesLength: In

which means we cannot link them with a function without casting: we know that the function to pass to transformOutput should have type String => Int, but the compiler doesn’t.

val readsLine: In =
  In.readLine
val computesLength: In
  In.transformOutput(str => str.asInstanceOf[String].length)

The key idea out of this problem is that we can add a type parameter to In which represents the type of its output. The resulting type In[A] lets us write:

val readsString: In[String]
val computesInt: In[Int]

Note that this doesn’t require us to actually perform any action, we’re still just building a datatype with a sealed trait and case classes, except this datatype now carries enough type information to allow for well typed composition. In other words, In[String] is not a container that contains a String, rather it’s a command to eventually read one, encoded as a datatype.

transformOutput can now have a proper type:

def transformOutput[A, B]: (In[A], A => B) => In[B]

This signature has two type variables (or type parameters), A and B . The rule with type variables is that whenever the same type variable is mentioned, the relative types have to match: in this case, (In[A], A => ... means that the input of the function needs to match the output of the In program, and ... => B) => In[B] means that the output of the resulting In program will match the output of the function. Therefore in the example above the function we need to pass to transformOutput to connect readsString: In[String] with computesInt: In[Int] has to have type String => Int, just like we expect.

Conversely, whenever different type variables appear, the relative types can be different, but they don’t have to, or in other words transformOutput also works if you use it with an In[String] and a String => String, resulting in another In[String].

We can now write a proper version of In:

/*
 * carrier:
 *   In[A]
 * introduction forms:
 *   readLine: In[String]
 * combinators:
 *   transformOutput[A, B]: (In[A], A => B) => In[B]
 */
sealed trait In[A] {
  def transformOutput(transform: A => B): In[B]
  ...
object In {
  val readLine: In[String]
  ...

and use it to express our original program:

val prog: In[Boolean] =
  readLine
    .transformOutput(input => input.length)
    .transformOutput(length => length > 10)

Finally, we need to complete In with an elimination form so that we can embed it into bigger programs, as usual we will translate to IO:

/*
 * carrier:
 *   In[A]
 * introduction forms:
 *   readLine: In[String]
 * combinators:
 *   transformOutput[A, B]: (In[A], A => B) => In[B]
 * elimination forms:
 *   run[A]: In[A] => IO[A]
 */
sealed trait In[A] {
  def transformOutput(transform: A => B): In[B]
  def run: IO[A]
  ...
object In {
  val readLine: In[String]
  ...

that being said, we won’t be thinking about eliminations forms for the next few articles, as we focus on writing programs with our algebras. We will return to the topic of elimination forms once we talk about IO in more detail.

Laws

You might be wondering why I have written the final program as:

val prog1: In[Boolean] =
  readLine
    .transformOutput(input => input.length)
    .transformOutput(length => length > 10)

as opposed to:

val prog2: In[Boolean] =
  readLine.transformOutput(input => input.length > 10)

prog2 seems less verbose, so should we refactor prog1 into prog2? Will the behaviour change? Intuitively, it would feel really weird if it did: transforming the output twice ought to be the same of transforming it once with the composite transformation.

We can encode this type of assumption as a law, something of shape:

expr1 <-> expr2

where <-> means that expr1 can be rewritten into expr2, and vice versa. In our case, we will say that:

p.transformOutput(f).transformOutput(g) <-> p.transformOutput(x => g(f(x)))

where:
  p: In[A]
  f: A => B
  g: B => C

which means that we can switch between prog1 and prog2 at will, and not just in the case where p = readLine, f = _.length, and g = _ > 10, but for any p, f, and g, as long they have the correct type. So in this case we use laws as a refactoring aid : they gave us freedom to refactor by specifying which transformations on our programs are harmless.

By the way, since Scala functions already have an andThen method to express function composition, the law above can be written as:

p.transformOutput(f).transformOutput(g) <-> p.transformOutput(f.andThen(g))

And as it turns out, there is another law concerning transformOutput, the fact that transforming an output with a function that doesn’t change it is the same as not transforming it at all:

p.transformOutput(x => x) <-> p

where:
  p: In[A]

If this seems completely obvious, that’s because it is! Many laws are just stating: my algebra behaves in the way you expect.

Conclusion

In this article we introduced a really important idea: encoding the output of a program by adding a type parameter to the carrier type of our algebra. This enabled us to add the transformOutput combinator, and next time we will use the same insight to model chaining, which is the essence of sequential control flow.