Foldable for non-Haskellers: Haskell's controversial FTP proposal

In the Haskell world, there is currently a big fuss about the “Foldable/Traversable in Prelude”-proposal.

Edit: For the record: it was a proposal, and has been implemented in the current version of Haskell (GHC)

As a non-Haskeller, you probably wonder what all the fuss is about.

First things first: some context

In most languages, you have functions to iterate over a sequence / array / iterable / list, for example in C# (assuming we have an array called values)

int[] values = new int[] {1, 2, 3};

foreach(var v in values) {
  Console.WriteLine(v);
}

In Haskell, we usually refer to an sequence with a list, and the example code would look like this:

values = [1,2,3]

mapM_ print values

Let’s say we need a function that concatenates several sequences of the type a.

In somewhat contrived C# we would write it like this (assuming we use arrays):

static Ta[] Concat<Ta>(Ta[][] values) {
    var res = new Ta[]{};
    foreach(var v in values) {
        res = res.Concat(v).ToArray();
    }
    return res;
}

int[][] values = new int[][] {
    new int[] {1, 2, 3},
    new int[] {4, 5, 6}
};

Concat(values);
// results in [1,2,3,4,5,6]

In Haskell we’d write it like this:

values :: [[Int]]
values = [[1,2,3],[4,5,6]]

concat :: [[a]] -> [a]
concat vals = foldr (\val acc -> val ++ acc) [] vals

concat values
-- results in [1,2,3,4,5,6]

Optional reading: type definitions

The odd someName :: a -> b -> c you see on top of the function implementations is a type definition.

  • The thing before the :: defines the name of the function
  • The last term defines the type of return value.
  • Other terms define type of the function’s parameters.

A few examples:

  • value :: Int
    • Takes no input
    • Returns a value of the type Int
    • Example: value = 5
  • values :: [Int]
    • Takes no input
    • Returns a list of values of the type Int
    • Example: values = [1,2,3]
  • values :: [[Int]]
    • Takes no input
    • Returns a list of a list of values of the type Int
    • Example: values = [[1,2,3],[4,5,6]]
  • decrement :: Int -> Int
    • Takes a parameter of the type Int
    • Returns a value of the type Int
    • Example: decrement x = x - 1
  • add :: Int -> Int -> Int
    • Takes two parameters of the type Int
    • Returns a value of the type Int
    • Example: addTwoInts x y = x + y
  • length :: [a] -> Int
    • Takes a list of a where a can be any type
    • Returns a value of the type Int
    • This function works on any list, no matter which is the type of a:
      • length [4,5,6] returns 3
      • length ['a','b','c'] also returns 3
      • length "abc" also returns 3, as a String is a synonym for a list of Char
  • decrement :: Num a => a -> a
    • Requires a type a to implement the Num type class.
      • A type class is somewhat similar to an interface in other languages.
      • The type class Num implements arithmetic operators for the type.
    • Takes a parameter of the type a
    • Returns a value of the type a
    • Example: decrement x = x - 1.

Now back to the main content…

What is a Foldable anyway

Notice that in the Haskell example the concat function is defined for a list of lists, i.e. concat :: [[a]] -> [a].

There is a well known functional paradigm called a fold, which allows you to convert something that has multiple values into a single value.

Let’s take the example of a sum:

  • Starts with an initial accumulator value of 0
  • Adds each of the sequence’s values one by one to the accumulator
  • returns the final accumulator.

This would be the code:

sum :: Num a => [a] -> a
sum vals = foldl (\acc val -> acc + val) 0 vals

product :: Num a => [a] -> a
product vals = foldl (\acc val -> acc * val) 1 vals

Now, imagine that you have a binary tree structure where a node contains either:

  • a node with a left and a right child node
  • a leaf with a value
  • an empty node:
data Tree a = Node (Tree a) (Tree a)
            | Leaf a
            | Empty

And that you need to get the sum and product for all these leaves. This goes as follows:

  • acc refers to the currently accumulated value
  • ln refers to a left child node
  • rn refers to a right child node
  • val refers to a value of a leaf node

If it’s a node, the XXXacc recursively calls itself to add or multiply it’s children’s values, if it’s a leaf, it just adds/multiplies the value with the accumulator.

sum :: Num a => Tree a -> a
sum tree = sumAcc tree 0
  where
    sumAcc node acc =
      case node of
        Leaf val   -> acc + val
        Node ln rn -> acc + (sumAcc ln) + (sumAcc rn)
        Empty      -> acc

prod :: Num a => Tree a -> a
prod tree = prodAcc tree 1
  where
    prodAcc node acc =
      case node of
        Leaf val   -> acc * val
        Node ln rn -> acc * (prodAcc ln) * (prodAcc rn)
        Empty      -> acc

You might see a pattern here. Haskellers hate copy paste programming, so they extract what is common.

If you take a look at what is common and different, you see only two differences:

  • We use + and *
  • We start from 0 and 1

So a Haskeller would extract this functionality, and as this is a common pattern, we’d call it a Foldable type class instance. Code would look like this

data Tree a = Node (Tree a) (Tree a)
            | Leaf a
            | Empty

instance Foldable Tree where
  -- some code here

sum :: Num a => Tree a -> a
sum tree = foldl (\acc val -> acc + val) 0 tree

prod :: Num a => Tree a -> a
prod tree = foldl (\acc val -> acc * val) 1 tree

Optional reading: cleaning up Haskell code

Even though this is compilable Haskell code, and the program would work, you would never see Haskell code like this in the wild. Haskellers apply some concepts to make the source code even leaner to read; let me show you how:

  • Let’s take a look at this expression:
    • sum tree = foldl (\acc val -> acc + val) 0 tree
  • The type declaration is sum :: Num a => Tree a -> a so we know the function
    • Requires a to implement the Num type class.
    • Takes a Tree that contains values of the type a
    • Returns a value of the type a
  • In Haskell you can “convert operators into functions” by putting parenthesis around them, so
    • \acc val -> acc + val is equivalent to \acc val -> (+) acc val
    • The type for this accumulator function is Num a => a -> a -> a
  • As both parameters now have the same order, we can apply a concept called currying (i.e. omit the last parameters if they are the same), so
    • \acc val -> (+) acc val is equivalent to \acc -> (+) acc.
    • The type for this accumulator function remains Num a => a -> a -> a
  • we curry it once more, so
    • \acc -> (+) acc is equivalent to \ -> (+).
    • The type for this accumulator function remains Num a => a -> a -> a
  • Calling a function with no parameters is the same as calling the function directly, so
    • \ -> (+) is equivalent to (+)
    • The type for this accumulator function remains Num a => a -> a -> a
  • Replacing this in the original declaration, we now we end up with
    • sum tree = foldl (+) 0 tree
    • The type declaration still is sum :: Num a => Tree a -> a
  • As the function’s parameters are similar once again, we can omit the tree part, so
    • sum tree = foldl (+) 0 tree is equivalent to sum = foldl (+) 0
    • The type declaration still is sum :: Num a => Tree a -> a

So a Haskeller would most likely end up with the following code, while maintaining the same types:

data Tree a = Node (Tree a) (Tree a)
           | Leaf a
           | Empty

instance Foldable Tree where
 -- some code here

sum :: Num a => Tree a -> a
sum = foldl (+) 0

prod :: Num a => Tree a -> a
prod = foldl (*) 1

This has one important consequence when you are doing Haskell:

In order to figure out what parameters a function requires, you need to look at it’s type declaration, and not at the implementation.

So both the sum and the product take a Tree of a, and return an a, where a implements the Num type class. This is the reason why Haskellers care so much about type definitions.

Back to the controversial FTP-proposal

Haskell works with a system they call modules (referred to as namespaces in other languages).

By default, Haskell applications import a module called Prelude, which contains a lot of handy helper functions, for example the sum function we mentioned in the beginning :

sum :: Num a => [a] -> a
sum = foldl (+) 0

Now, as Haskellers do not only want to use sum on a list, but also on any Foldable instance, they decided to change the type signature to this:

sum :: (Foldable t, Num a) => t a -> a
sum = foldl (+) 0

Because there exists a Foldable instance for a list, we can use the “Foldable sum” function, and remove the sum function that only works on lists.

So what is controversial about this proposal?

People come from different backgrounds. When you talk about a list, people new in Haskell think about a sequence, and they can imagine what is happening.

However, due to the new signature (i.e. the Foldable thing), people new to Haskell might have a hard time to get started with Haskell, because they need to learn about the concept of Foldable first, so the barrier of entry is getting larger.

In my opinion, it might be a little harder to grasp at first, but if, as a beginner, you are willing to just accept a few things without knowing why they are like they are, the barrier to entry shouldn’t be much higher than before.

UPDATE

I assumed I understood what all the fuss was about, but apparently I did not. Luckily, @bitemyapp went the extra mile to explain while it was wrong:

I’d suggest you to read his article, but I’ll give you the recap here:

Next to a Foldable, we also have a Functor. A Functor is something that you can map over, a structure that contains zero, one or more values. An example in GHCI:

Prelude> let inc = (1 +)
Prelude> inc 5
6
Prelude> fmap inc [1,2,3]
[2,3,4]
Prelude> fmap inc (Just 5)
Just 6
Prelude> fmap inc Nothing
Nothing
Prelude> fmap inc (Right 4)
Right 5
Prelude> fmap inc (Left 3)
Left 3

Just think of this as a map or select (C#) statement. In the example above we have the following statements:

  • let inc = (1 +)
    • is of the type inc :: Num a => a -> a, so it takes a Num instance, and adds 1 to it?
  • inc 5
    • returns 6
  • fmap inc [1,2,3]
    • returns [2,3,4]
    • the container is a List, and the values in the list are Num instances.
    • the type signature for fmap is : fmap (a->b) -> [a] -> [b]
    • fmap takes a transformation function and a list, and returns a List containing the transformed values in the same order.
    • For a list, Haskell’s fmap is equivalent to map, which is equivalent to a C# select statement.
    • In C#, fmap could be implemented for a list as List<Tb> fmap<Ta,Tb>(Func<Ta,Tb> transform,List<Ta> list) => list.Select(transform).ToList().
    • Bifunctors:
      • The Maybe type is a type that either represents a value or nothing.
        • this type was introduced to avoid null checking in all of your code.
        • values are represented using Just a where a can be any value.
        • nothing is represented using Nothing.
        • when you want to change the value of a Maybe, you typically only want to change it, when it contains a value. If it does not contain a value, you just want to do nothing with it.
        • how do we change the value of something in a container when using Haskell? Exactly: using fmap
        • fmap inc (Just 5)
          • returns Just 6, so it applied the inc function to the value of the Maybe
        • fmap inc Nothing
          • returns Nothing, because it doesn’t make any sense to apply changes to a “null value”
        • let’s say we have a var called maybeVal, that has the type Maybe Num, so it is either Nothing or Just x where x is a number.
        • if we want to either increment the value if it’s defined, or just return a 0 if it is not, we could do it like this:
        • maybe 0 inc maybeVal
          • if maybeVal is Nothing it uses the first argument : 0
            • maybe 0 inc Nothing returns 0
          • if maybeVal has a value (Just 5), it applies the inc function to the value and returns that value, i.e. 6
            • maybe 0 inc (Just 5) returns 6
          • fmap is the Functor implementation, maybe is the Bifunctor implementation.
      • the Either type
        • Now, avoiding null checks everywhere is fun and all, but what if you would like to return some kind of error message in case a value is not defined?
        • For this we have the Either a b type, that can return 2 different types:
          • It can Left a, where Left a is typically an error message
          • It can return a Right b, where Right b is similar to a Just b, so this contains the value.
          • Let’s say we have a function to parse a date from a string. We want it to return either the date, or an error message.
            • this would be the type of the function: parseDateFromString :: String -> Either Error Date
            • this would be equivalent to Either<Error,Date> parseDateFromString(String)
            • in case there is a problem with the parsing we’d get return value like this: Left (Error "Invalid date format; expected YYYY/MM/DD")
            • in case there is no problem with the parsing we’d get the value like this: Right (Date 2015 10 20)
          • What if we would like to print the year if it’s a valid year, or display the error when it is not? Let’s assume we have a value eitherErrorOrDate
            • we would call either displayError printYear eitherErrorOrDate
              • the type of eitherErrorOrDate would be Either Error Date
              • for now, just take my word for it: a type of IO () means: something that interacted with the outside world, but didn't return any value
              • the type of the function displayError would be Error -> IO (), so it takes an error, and implies a changed world when executed.
              • the type of the function printYear would be Date -> IO (), so it takes an error, and implies a changed world when executed.

Now what’s wrong with the controversial FTP proposal:

In a Maybe or an Either, it makes sense that you only want to map to the Just value or the Right value.

Now, let’s say you have a Tuple, and would map over this:

  • A tuple has the type (a,b)
  • So when you fmap inc over (1,2), you’d expect a (2,3), correct?
  • Well, you don’t : GHCI returns a (1,3)
  • This is because a tuple can contain 2 different types, f.e. ("Hello",4)
  • As an fmap can only take one type as an input parameter, they decided to go for the second element, so fmap inc ("Hello",4) returns ("Hello", 5).
  • This doesn’t make sense, as the first element might be just as important in a tuple as the other, so this shouldn’t have been a Functor in the first place.

Thank you Chris for taking the time to explain this!