Intro to type classes

Have you ever wondered what is the essence of object-oriented programming? Some people will tell you that object orientation is a non-concept plagued by epistemological problems while most of the people will pick different, sometimes incompatible things.

As Paul Graham points out, the vague concept of OO is typically a subset of encapsulation, protection, several kinds of polymorphism, everything-is-an-object-ness, message passing, inheritance, etc. In his ironic words:

Because OO is a moving target, OO zealots will choose some subset of this menu by whim and then use it to try to convince you that you are a loser.

This very old quote of Alan Perlis can be understood as a critique of the object-oriented style:

It is better to have 100 functions operate on one data structure than 10 functions on 10 data structures.

In fact, one problem of having rich object models with many classes is that we lose the leverage that many functions operating on a single type could bring us. The standard OO tool for this problem is the polymorphism. A single method call (declared in an interface or by duck typing) can be directed to many different implementations. In this spirit, Rich Hickey updated the old Perlis idea (emphasis mine):

It is better to have 100 functions operate on one data abstraction than 10 functions on 10 data structures.

A magnificent example of this concept is the collections library of his own programming language, Clojure. With a couple of functions from the about 100 functions summarized on Clojure’s documentation site, you can express almost any collection transformation function you ever need to write1.

However, doing this with classical polymorphism is only easy when you control the code of all the classes that need to implement the interface. Hickey’s can easily make sets to comply with ISeq but you might be frustrated trying to get a value from a library by creating subclasses (tricky and no always possible) or the adapter pattern2 (cumbersome and sometimes impractical).

But there is a technique addressing these problems (and the more general expression problem) called ad-hoc polymorphism. Some programming languages implement different flavors of it either directly or as a design pattern.

Motivating example in Haskell

Let’s start with a Haskell example, as it is the programming language that has popularized type classes3 as a form of ad-hoc polymorphism. Imagine that you need to convert a list of strings into a JSON array.

toJson1 :: [String] -> String
toJson1 elems = toArray (map quote elems)
    where quote elem = '"' : elem ++ "\""
          toArray elems = '[' : concat (intersperse ", " elems) ++ "]"

This code works as expected:

*Main> toJson1 ["a", "b", "c"]
"[\"a\", \"b\", \"c\"]"

What happens when we want to generalize the input from [String] to a generic list? Most OO programming languages have a toString or to_s in the top type4 so you can always convert to string whatever you receive. However, in Haskell there is no inheritance or subtype relations and so no Object.toString to call, but there is a show function. Let’s use it and let the compiler infer the function signature.

toJson2 elems = toArray $ map show elems
    where toArray elems = '[' : concat (intersperse ", " elems) ++ "]"
*Main> toJson2 ["a", "b", "c"]
"[\"a\", \"b\", \"c\"]"
*Main> toJson2 [1, 2, 3]
"[1, 2, 3]"
*Main> :t toJson2
toJson2 :: Show a => [a] -> [Char]

Now it works both for string and numbers! More interestingly, the Show a => part of the signature means that this function works with any type having an instance of the Show type class. Most standard types are show-ables but functions are not.

*Main> toJson2 [(+), (-)]
<interactive>:23:1: error:
    • No instance for (Show (a0 -> a0 -> a0))

To make a new type show-able we just need to instantiate the class for it. The relevant part of the Show class is:

class Show a where
  show :: a -> String

We can introduce some domain specific type and give it a show instance:

data Conversation = Conversation { from :: Int
                                 , to   :: Int
                                 , text :: String }

instance Show Conversation where
    show (Conversation from to text) =
      concat [ "{\"from\":", show from, ",",
                "\"to\":", show to, ",",
                "\"text\":", show text, "}" ]

And it works like a charm!

*Main> putStrLn $ toJson2 [Conversation 1 2 "hi!", Conversation 2 1 "hello"]
[{"from":1,"to":2,"text":"hi!"}, {"from":2,"to":1,"text":"hello"}]
*Main> toJson2 [Conversation 1 2 "hi!", Conversation 2 1 "hello"]
"[{\"from\":1,\"to\":2,\"text\":\"hi!\"}, {\"from\":2,\"to\":1,\"text\":\"hello\"}]"

However, we are misusing the Show instance. Its purpose is having canonical string representation of data (that you might later parse with Read) and, for most types, is not what we want. By the way, you can get a Show instance for free by using the deriving keyword.

data Conversation = Conversation { from :: Int
                                  , to :: Int
                                  , text :: String
                                  } deriving (Show)
*Main> show (Conversation 1 2 "hello world")
"Conversation {from = 1, to = 2, text = \"hello world\"}"

Our own class

But nothing precludes us from creating a EncodeJson type class instead of relying on Show.

class EncodeJson a where
  toJson :: a -> String

The instances from integers and strings will delegate to show and the instance for Conversation can be very similar to the earlier Show Conversation. For other types will be able to have arbitrary implementations that will be called polymorphically regardless of having no inheritance mechanism.

At this point we can revisit our original function and use EncodeJson instead of Show.

toJson3 :: EncodeJson a => [a] -> String
toJson3 elems = toArray $ map toJson elems
    where toArray elems = '[' : concat (intersperse ", " elems) ++ "]"

Composing type classes

If you look carefully at toJson3 you will find that it is exactly what you would expect from the implementation of EncodeJson for a list of JSON-able things. This is a very powerful pattern in which instances of more complex types are created from instances of simpler ones. You just need to use a type variable with a bound using the same syntax that in the function (EncodeJson a =>):

instance EncodeJson a => EncodeJson [a] where
    toJson elems = toArray $ map toJson elems
        where toArray elems = '[' : concat (intersperse ", " elems) ++ "]"

This is very powerful since we can uniformly use toJson for anything we can combine from what we have defined. For example, a single string, lists of lists of strings or a list of conversations.

*Main> toJson "hi"
*Main> toJson [["a", "b"], ["a"]]
"[[\"a\", \"b\"], [\"a\"]]"
*Main> toJson [Conversation 1 2 "hi", Conversation 2 1 "hello"]
"[{\"from\":1,\"to\":2,\"text\":\"hi\"}, {\"from\":2,\"to\":1,\"text\":\"hello\"}]"

Flexibility and separation of concerns

One notable feature of type classes is that you can define class instances for types defined in other modules and for types you have no control at all. You might write a EncodeJson for a type from a third-party calendar library that knows nothing about your API format.

You might want to have instances defined in a separate module to foster separation of concerns. For example, you might have a core module with types that model the domain and two other different modules defining how to map them to JSON and how to persist them.

Parting thoughts

Type classes are a powerful tool to build composable software, specially in languages that implements them like Haskell or Rust. They are also extremely useful in Scala despite being implemented as a design pattern on top of implicits. They are more verbose but, believe it or not, more flexible than Haskell’s ones in some respects.

We will take a look at type classes in Scala in an upcoming article. Meanwhile, you can see the complete code of this one in this gist.

  1. I realized that by solving most problems in 4clojure that way. 

  2. Look mum! A GoF pattern other than singleton! 

  3. class in type class has nothing to do with object oriented classes. It should be understood as a family of types that can place a role. 

  4. this base class is usually a dumpster of type unsafe stuff, like Java’s comparison operator that allows you to compare String against Int as is it made sense.