Explain Free Monad Like I am Five (Part 2)

Photo by Esther Jiao

In part one of this post, we talked about how Free Monad can build a monad from Functor. To do so, you can create a DSL like programs and generate another program. The program that you make becomes plain data structures that represent a domain logic. Once you composed the program, you can connect those programs with an interpreter. You can switch a different interpreter on dev and QA. You can also abstract out the implementation of your program or optimize them without changing your domain logic.

We started by creating a Todo application with these algebras:

case class Todo(id: Long, description: String, isFinished: Boolean)

sealed trait Action[T] extends Product with Serializable
case class Create(description: String) extends Action[Todo]
case class Find(description: String) extends Action[Option[Todo]]
case class Mark(id: Long) extends Action[Unit]
case class Delete(id: Long) extends Action[Option[Todo]]
case object Read extends Action[List[Todo]]

First, we created an algebra FlatMap and Pure and introduce what it’s like to make a monad like programs by wrapping our Todo algebras into the “Free” data structure:

sealed trait Free[F[_], A] {
  import Free._

  def flatMap[B](fn: A => Free[F, B]): Free[F, B] = this match {
    case Free.Pure(a) =>
      fn(a)
    case Free.FlatMap(fa, fun) =>
      FlatMap(fa, fun andThen (a => a.flatMap(fn)))
  }
  def map[B](fn: A => B): Free[F, B] = flatMap(a => Pure(fn(a)))
}

object Free {
  final case class FlatMap[A, B, F[_]](fa: F[A], fun: A => Free[F, B]) extends Free[F, B]
  final case class Pure[F[_], A](a: A) extends Free[F, A]

}

Then, we created a way to “lift” these algebras to a “Free” monad by creating an implicit conversion of Action[A] => Free[Action, A]:

implicit def lift[A](fa: Action[A]): Free[Action, A] = FlatMap(fa, Pure.apply)

Lastly, we ended with connecting the program and the interpreter with our “Free” algebras that we construct.

def runProgram[A](program: Free[Action, A]): A = program match {
    case Free.Pure(a) => a
    case FlatMap(fa: Action[A], fn: (A => Free[Action, A])) =>
      // execute the Action here

      val res = execute(fa)
      // thread the function into a new Free
      val newFree = fn(res)
      // execute the next function
      runProgram(newFree)
  }

Note: This is the short recap from “Explain Free Monad like I’m Five (Part one)” I explain my thought process on how I came up to these conclusions. If you are interested, check it out before reading further.

Problem

The above solutions seem to work for constructing a program using Free Monad for a Todo Application. However, that solution will only work for that particular application. We cannot use our runProgram function or “lift” our algebras to a “Free” version. It will be quite tedious if we need to implement the runProgram for every program that we construct and also creating an implicit conversion each time we build a new set of algebras.

Generalizing Lift

To generalized lift, we just need to abstract out Action to an F[_] type constructor:

implicit def lift[F[_], A](fa: F[A]): Free[F, A] = FlatMap(fa, Pure.apply)

However, if we do this, it will cost compile error because a lot of the times, we will want to lift concrete type constructors instead of the generic “sealed trait”:

val create: Free[Create, String] = lift(Create("Check Homework"))

This will throw an error in for-comprehension because it cannot automatically lift the value to the Action sealed trait.

One way to do this is to create a “smart-constructor”, by first setting the value as the generic algebra Action, and defined the function that abstract out regular input to a “free” version of it.

type TodoAction[A] = Free[Action, A]

def create(description: String): TodoAction[Todo] = lift(Create(description))
def find(description: String): TodoAction[Option[Todo]] = lift(Find(description))
def mark(id: Long): TodoAction[Unit] = lift(Mark(id))
def delete(id: Long): TodoAction[Option[Todo]] = lift(Delete(id))
def read: TodoAction[List[Todo]] = lift(Read)

Remember that the return of the “Free” type is defined by the generalized algebra return type. For instance, create will return a TodoAction[Todo] because the algebra for Create returns an Action[Todo]:

case class Create(description: String) extends Action[Todo]

Then we can construct our program with these smart constructors:

val program =
  for {
    todo <- create("Do Laundry")
    _ <- mark(todo.id)
    listOfTodo <- read
  } yield {
    println(listOfTodo)
  }

Generalizing RunProgram

Referencing from the runProgram from the previous article:

def runProgram[A](program: Free[Action, A]): A = program match {
    case Free.Pure(a) => a
    case FlatMap(fa: Action[A], fn: (A => Free[Action, A])) =>
      // execute the Action here

      val res = execute(fa)
      // thread the function into a new Free
      val newFree = fn(res)
      // execute the next function
      runProgram(newFree)
  }

The current run program has an execute function tied to the ones with Action. Moreover, Free[Action,A] is restricted to the particular Todo application. How can we generalize this further to take any program of F[_]?

We can generalize the runProgram by taking a type constructor F[_], but how will we generalize the execute function?

def runProgram[F[_], A](program: Free[F, A]): A = program match {
  case Free.Pure(a) => a
  case FlatMap(f: F[A], fn: (A => Free[F,A])) =>
    val res = execute(fa) // where is this execute come from?
    val newFree = fn(res)
    runProgram(newFree)
}

Somehow, the execute function should be a generic function that the user defines when creating the program. Therefore, it needs to be some trait that will need to pass in as an argument. If you look closely at the implementation above, the function execute(fa) takes in an F[A] and returns an A. That A, then, can be an input to the fn function. Therefore, we need to define the execute function that takes an F[A] and returns an A. Let’s define the execute function inside an Executor trait:

trait Executor[F[_]] {
  def execute[A](fa: F[A]): A
}

We pass in the Executor trait class as an argument in runProgram, and call the execute within that Executor trait:

def runProgram[F[_], A](program: Free[F, A], executor: Executor[F]): A = program match {
  case Pure(a) => a
  case FlatMap(fa: F[A], f: (A => Free[F, A])) =>
    val res = executor.execute(fa)
    runProgram(f(res), executor)
}

Then, we can define our interpreter for Todo application in the Executor, execute function and pass the value into the runProgram:

runProgram(program, executor)

We finally can run Free[F,A] on any programs by providing the custom “execute” that interprets our algebra. However, can we generalized this even further?

Generalizing Executor

The current Executor function takes in an F[A] and returns an A. However, this function is limited to the pure effect type. We can somehow generalize this further by generalizing its effect type. If you are unsure about what effect means, check out my article on What is “effect” or “effectful” mean in Functional Programming? | edward-huang.com.

This kind of transformation between functors is called natural transformation. Instead of having a signature of F[A] => A, which is equivalent to F[A] => Id[A], we can have F[A] => G[A] where G[_] can be any type constructor.

Let’s refactor our Executor function to take in 2 type constructor:

trait Executor[F[_], G[_]] {
  def execute[A](fa: F[A]): G[A]
}

Then, runProgram will take an additional type constructor G[_]:

def runProgram[F[_], G[_], A](program: Free[F, A], executor: Executor[F, G]): G[A] = program match {
  case Pure(a) => a
  case FlatMap(fa: F[A], f: (A => Free[F, A])) =>
    val res = executor.execute(fa)
    runProgram(f(res), executor) // something is not right here ....
}

You noticed that the above code would not compile. Initially, execute function returns A. However, after changing our Executor function to return G[A], we cannot pass in the result to the f function in FlatMap.

We can resolve this issue by requiring G to be a Monad, and use flatMap to bind the value inside of G to f:

import cats.{Monad}

 def runProgram[F[_], A, G[_]](program: Free[F, A], executor: Executor[F, G])(implicit M: Monad[G]): G[A] =
    program match {
      case Pure(a) => M.pure(a)
      case FlatMap(fa: F[A], f: (A => Free[F, A])) =>
        val res = executor.execute(fa)
        M.flatMap(res)(a => runProgram(f(a), executor))
  }

Conclusion

In this article, we learn how to generalize our application even further by generalizing our “Free” wrapper to use it in various applications.

First, we generalize the lift functions and introduce a “smart-constructor” to construct our program. Then, we generalized our runProgram, the interpreter that runs the program, by submitting an Executor trait. Lastly, we generalized our Executor trait further by introducing natural-transformation on our Executor trait.

The Free structure exists in the Cats library with different names. Hopefully, you can understand the concept of Free and how Free works. There are also Free Applicative, and the difference is in the runProgram and its “Free” algebra, where instead of being a Monad, it is an Applicative functionality.

All source code on this article can be viewed here.

Source and Reference: Free Monads Explained (pt 1). Building composable DSLs | by Oleksii Avramenko | Medium Cats: FreeMonads

Like this Article?

Sign up for my newsletter to get notified for new articles!


Related Posts

5 Anti Pattern for Writing Code in a Functional Programming Language

No 1. Nested Asynchronous Function

Why Do Functional Programmers Prefer For-Comprehension Over Imperative Code Block

Short Answer - Readability

How to Turn Domain Model into DynamoDB AttributeValue

A brief introduction about Dynosaur

Why is REST more popular than RPC?

REST is much more flexible

Want to Safe more money and Resources for your Company? Treat Your Software as a Living Organism

Codebase are not machines but a living organism