Why We Need Monads
Monads are often a point of frustration for learners of Haskell. The question of what a monad is leads to the question of why a monad is. The question of why a monad is leads to the question of what a monad is. It's been said that monads aren't necessary. This is wrong. It's time that these myths be debunked. This article seeks to answer the question of why a monad is.
The Basics of Haskell
Haskell allows one to write very elegant programs. Let’s begin by looking at a program that calculates the factorial of it’s input.
Here we have a function of 1 argument. The function’s definition is a single if then else
statement. I don’t want to dwell on this example for too long. It should be clear what’s going on in this example by inspection. The important thing to note is how functions and if-statements are defined.
Let’s look at a slightly more involved example that calculates the hypotenuse of it’s 2 inputs. In this function we’ll see an example of multi-line comments, a function of more than 1 argument and, the let-in clause.
Notice at the very top of this program that we have {- -}
. These are how one would write multi-line comments. Later we’ll see an example of single-line comments.
The last function had a single argument. This function has 2 arguments. Note the syntax for that. The syntax takes the form f arg1 arg2 arg3 ...
where the trailing ellipsis could be additional arguments.
The body of this function is also quite different from our first example. Here we’re using a let-in clause. We’re saying let a'
be a2 and, let b'
be b2. Once we’ve defined a'
and b'
we can use them in an expression. Here the expression is sqrt (a' + b')
.
These small examples should feel straight forward. The next example will be far more involved. We’ll be using a bisection search to find the square root of some number x. This example will use a function defined locally within another function, single-line comments show up, the let-in and if-statements make a return and, a new concept called a where clause will be introduced.
{-
Find the sqrt of x.
-}
sqrt' x =
-- The sqrt of x must be between 0 and x
-- so, that will be our initial guess.
guess 0 x
where
-- This is the tolerance.
-- The program will find the sqrt +/- epsilon.
epsilon = 0.001
guess low high =
let
rt = (high + low) / 2 -- Take the midpoint as the potential square root.
x' = rt ** 2 -- Squaring the guess should be the original value x.
in
if abs (x' - x) >= epsilon
then
if x' < x
then guess rt high
else guess low rt
else
rt
First take a look at these comments. At the top we have a multi-line comment just as before but, now we have single-line comments denoted by --
. Hopefully, the syntax highlighting makes these comments clear to distinguish.
Look at the body of the sqrt'
function. It’s just a single statement. Our function’s body is just a single function call in the form guess 0 x
. guess
is a function defined within a where clause. In fact, we define 2 things in the where clause. We defined guess
and, a variable called epsilon
. You may be thinking that a where clause and a let-in clause are very similar. Indeed, they can be thought of as doing the same thing in a different way.
I’ll leave the analysis of the guess
function as an exercise to the reader. Syntactically speaking guess
contains no new language constructs however, that doesn’t mean understanding the program is trivial. Much has been written about this method. In fact, the bisection search of square roots have been known since Babylonian times!
So far we’ve seen quite a bit about Haskell. We’ve seen ways to define functions, if-statements, the let-in clause and, the where clause. One big topic I’m not going over is algebraic data types and pattern matching. Make no mistake. These are extremely important topics but, for the sake of brevity I won’t go over those here.
The Problem of Haskell
In general there are 2 ways of evaluating programs. The most common evaluation strategy is called applicative order evaluation. The other evaluation strategy is called normal order evaluation. Haskell uses normal order evaluation.
Unfortunately normal order evaluation has a very nasty problem. That problem is side effects. A pure function is strictly a mapping from it’s domain to it’s range. If we want to write to stdout we don’t really care about it’s mapping. We want a function that has the side effect of writing to stdout. That causes a problem for a programming language like Haskell. Let’s look at a small example using applicative order evaluation first to get a better idea of everything.
This is our setup. numWithPrint
takes in a number, prints it to stdout, and returns the same number. So, numWithPrint 6 = 6
, numWithPrint 1 = 1
, etc. doubleNum
just doubles it’s input. E.g. doubleNum 4 = 8
. This is not valid Haskell code. We’ll find out why that is soon enough but, for now let’s use applicative order evaluation to evaluate doubleNum (numWithPrint 6)
.
1> doubleNum (numWithPrint 6)
# We evaluate the arguments of `doubleNum` first
2> doubleNum (print 6; 6)
# We can't reduce `doubleNum`'s arguments anymore
3> doubleNum 6
4> 6 + 6
5> 12
---
stdout:
6
Do you see what’s happening here? We evaluate the function’s arguments first and, substitute the result into the function’s body. doubleNum (numWithPrint 6)
got reduced to doubleNum 6
but, we have a side effect. The side effect was printing to stdout. Let’s look at what happens if we used normal order evaluation.
1> doubleNum (numWithPrint 6)
# We substitute the arguments of `doubleNum` first
2> (numWithPrint 6) + (numWithPrint 6)
# We can't substitute anymore
3> (print 6; 6) + (print 6; 6)
4> 6 + 6
5> 12
---
stdout:
6
6
Oh my! This isn’t good. We’ve printed ‘6’ twice! Unlike applicative order we don’t evaluate the function’s arguments first. Instead we substitute the arguments into the function’s body. This is the problem of Haskell.
Haskell’s solution to this problem is to use monads. Actually I want to make this very clear. Monads solve the problem of normal order evaluation. If you’re reading this paper surely at some point you’ll read that monads allow for pure functional programming. Not all functions in Haskell are pure functions. You may read that monads represent a type that passes along a world concept or, that they’re a type that represents the abstract idea of a computation. This is true at a conceptual level. You may read that monads aren’t necessary. For handling side effects in Haskell they are necessary. Let’s look at why they’re necessary with a small example of how to use them.
numWithPrint num = do
print num
return num
doubleNum ioNum = do
num <- ioNum
return (num + num)
doubleNum (numWithPrint 6)
The first thing that one needs to understand is the do-return macro. Look at numWithPrint
after the =
there’s a do
. This begins the do-return macro. The last line of this function is a return
. What are we returning? A monad! No longer is numWithPrint 6 = 6
a true statement. Rather numWithPrint 6 = IO 6
is what we have here.
This is quite good because doubleNum
only accepts a number wrapped in a monad. Unfortunately, we can’t just add 2 monads. It doesn’t work that way. We need to unwrap our monad and add them. That’s what num <- ioNum
does. It takes the monad ioNum
and, binds it’s value to num
. Now num
is just a regular number. We can double it and, return it back out as a monad.
This is valid haskell code. It isn’t as nice looking as before but, it’ll get the job done. Let’s look at this using normal order evaluation. There is one caveat about this analysis. You may have noticed I called do-return a macro. It’s useful to see what do-return expands to however, I won’t be expanding the do-return macro here.
1> doubleNum (numWithPrint 6)
2> doubleNum (print 6; return 6)
# Binding acts like a wall. We can't substitute
# anything at this point without evaluating `num`
3> num <- (print 6; return 6);
return (num + num);
# 6 will be bound to `num`
4> num <- IO 6
return (num + num);
5> return (6 + 6)
6> IO 12
---
stdout:
6
Monads save the day. We can have our normal order evaluation and use it too. This is why we need monads.