It's said that all Haskell programmers eventually write a blog post explaining Monads. Today I fulfill my destiny. The basic format is ripped from "Learn You A Haskell For Great Good," but these are my own crappy examples.
This blog post is literate Haskell; you can get it here. You can import it into ghci and type the functions to see their output (
ghci monads.lhs). For this to work, we'll need to import one module for use later:
A functor generalizes the notion of types to a type with context. For example, the Maybe type is literally this:
Maybe a = Just a | Nothing
It wraps a simple type into the notion of uncertainty, so Maybe is a functor. The list type is really
List a is a list, but you can think of lists as functors generalizing types into the notion of non-determinism. To be a functor, a type must implement
fmap which has a type signature of
(a -> b) -> f a -> f b. Or concretely:
example1 = fmap (*2) (Just 4)
An applicative functor takes this further and generalizes the notion of function application inside a context. In addition to
fmap, a type that wants to be an applicative functor must implement
<*> :: (a -> b) -> f a -> f b. The only difference between it and
fmap is that the first argument, the function to be applied, must be inside the context already. Take a breath and read this example:
even? x | x `mod` 2 == 0 = True | otherwise = False
example2 = (even?) <$> [1..12] -- <$> is just fmap, but sexier to go with <*>
Note that Haskell gives Lists special treatment because they're so fundamental. So we write
[a] instead of
List a and, though we didn't show it here,
fmap is just the normal
map function you expect.
example3 = (*) <$> [1,2] <*> [3,4,5]
The result of applying
1 is a function that accepts one argument and returns the product.
<$> will return this new function wrapped in the List context and
<*> can then apply this contextualized function to
[3,4,5], in order.
example4 = [ (*1), (*2) ] <*> [1..4]
Note the length of the resulting lists is the product of the length of each list. One more to drive that point home:
example5 = [ (*x) | x <- [1..3] ] <*> [1..4]
Where functors generalize types inside contexts, and applicative functors generalize application inside contexts, monads generalize composition inside contexts.
A monad must implement
>>= :: m a -> (a -> mb) -> m b, which you pronounce as "bind." It takes a value wrapped in a monad context (
m) and a function from a non-monadic value to a monadic value, and applies them. The result is something that can be used all over again in the same way. Example!
example6 = ((take 5) . repeat) 2 -- signature is just a -> [a]
example6 returns a List of values, so it returns a monad.
example7 = [2,3] >>= ((take 5) . repeat)
Haskell has a special notation for monad computations called do-notation. It should illuminate the relationship between imperative programming and monads:
example8 = do l <- [2,3] ((take 5) . repeat) l
But let's focus on what just happened: I took a list of values and a function to act on a single value and applied it to the entire list without explicit iteration.
That's the magic of monads, really. A monad is a polymorphic type that generalizes the notion of function composition, and because of this you can do interesting things during composition. The Maybe monad does similar things:
ensureEven x | x `mod` 2 == 0 = Just x | otherwise = Nothing
example9 = do a <- ensureEven 2 b <- ensureEven $ a * 2; return b
example10 = do a <- ensureEven 2 b <- ensureEven $ a + 1 return b
So, a summary:
- Functors generalize types;
- Applicative functors generalize application;
- Monads generalize composition.
The point of generalization is to let you assign some interesting or novel meaning to these abstract ideas. List models non-determinism; Maybe models simple uncertainty; IO models, uh, IO actions.
That's why it kind of misses the point to say "monads model state." Certainly they can, but what they really do is so much more powerful. Here is
example10 without the do-notation:
example11 = (ensureEven 2) >>= (\a -> ensureEven (a + 1) >>= (\b -> return b))
The result of
ensureEven 2 is fed into a single function: an anonymous function with a free variable
a. It in turn contains an inner anonymous function
(\b -> return b), which, thanks to the magic of lexical scope, can see
a. And that's why
example12 = do putStrLn "Enter some text:" input <- getLine return input
can safely perform IO: the IO monad generalizes composition such that functions are called in the order they need to be in the real world (ie, asking for text before listening for it) and the "variables" can't be shared or mutated without you knowing.
Hope this made some sense.