‹ home

Dabbling in Erlang, part 2: A minimal introduction

12 Oct 2013

It’s been a while since the last post which was a short introduction into functional programming in general and now it’s time to dive into some of the basic concepts of the language.

Single assignment & pattern matching

In Erlang we should think variables in a mathematical sense of the term, which means that once you’ve bound a variable, you cannot change its value . If you state that X is 5, then it will always be 5 (a variable is called bound if already contains a value, unbound otherwise):

Eshell V5.10.1  (abort with ^G)
> X = 5.
> X = 6.
** exception error: no match of right hand side value 6
> X = 5.

But this error is not exactly what we expected. In Erlang, the = operator does actually pattern matching (there are no LHS/RHS values). A pattern match is done like so:

Pattern = Expression

The Pattern consists of data structures that may contain both bound and unbound variables. The Expression part may contain data structures, bound variables, mathematical operations and function calls. If the pattern match succeeds, any unbound variables will be bound and the value of the expression will be returned. Using pattern matching we can extract data out of complex structures:

> {_, Name, Surname} = {person, "John", "Doe"}.
> Name.
> Surname.
> {_, Name, "bla"} = {person, "John", "Doe"}.
** exception error: no match of right hand side value {person,"John","Doe"}

We extract the name and surname and bound them to the Name and Surname variables. The second pattern match fails because “bla” does not equal to “Doe”.

Note that all variable names must begin with a capital letter.

But this isn’t the end. Using pattern matching we can control the execution flow of our programs and write more conscise code. Let’s say we want a function which, depending on the first argument we pass it, will print something to the screen.

Here’s how we might do this in Ruby:

def greet(inc)
  case inc
  when "Hello"   then "Hi"
  when "Goodbye" then "See you"
  else inc

greet("Hello")   # => "Hi"
greet("Goodbye") # => "See you"
greet("blabla")  # => "blabla"

Here’s a direct port to Erlang:

greet(Inc) ->
  case Inc of
    "Hello"   -> "Hi";
    "Goodbye" -> "See you";
    _         -> Inc

We essentially use the case syntax which does pattern matching. But that’s not idiomatic Erlang; we can do better: pattern match in the head of the function clause (the part before the -> is called the head and after it comes the body).

greet("Hello")   -> "Hi";
greet("See you") -> "See you";
greet(Other)     -> Other.

A function is defined as a collection of clauses. Each clause specifies the expected argument patterns and a body of expressions to be evaluated. greet/1 is still one function but without the use of case or if statements. The name/arity notation is used to describe functions. So greet/1 means the greet function which accepts one argument. It’s worth mentioning that when the compiler creates a binary search tree with the greet clauses, so pattern matching is very efficient.

But it gets better.

Let’s say you want to calculate the area of a shape which could be either a square or a circle.

In Ruby we could define two different functions:

def area_square(side)
  side * side

def area_circle(radius)
  3.14 * radius * radius

area_square(15) # => 225
area_circle(5)  # => 78.5

or using a single function we might do:

def area(shape, x)
  if shape == :square
    x * x
  elsif shape == :circle
    3.14 * x * x
    raise "Unknown shape"

area(:square, 15) # => 225
area(:circle, 5)  # => 78.5

In Erlang we could do the following, using pattern matching and one function:

area({circle, Radius}) -> 3.14 * Radius * Radius;
area({square, Side})   -> Side * Side.

and we would use it like this:

> area({square, 15}).
> area({circle, 5}).
> area({wtf, 5}).
** exception error: no function clause matching area({wtf,5})

In the Erlang version, we’re defining one function with multiple clauses. We also don’t have to explicitly raise an error as we did in the Ruby version, since pattern matching will do this for us.

Note that we’re using a composite data type called a tuple (that thing inside the curly brackets). I won’t get into details here, but you will see tuples everywhere in Erlang.

It’s important to point out that a function is identified by its name and its arity. Two functions with the same name but different arity are two completely different functions that just happen to share the same name. Think of it like two people with the same first name, but different last names (ie. arity).

This means we can do this:

adder(_)       -> "I'm someone".
adder(_, _)    -> "I'm another one".
adder(_, _, _) -> "I'm yet another one".
> adder(3,2)
"I'm another one"
> adder(1)
"I'm someone"
> adder(5,6,7)
"I'm yet another one"

You can notice that we terminate each definition with a period, which means we are essentially defining a different function every time. _, like in Ruby, denotes that we don’t care about the actual values of the arguments.


Guards can be thought of as additional constraints that can be applied when pattern matching.

Let’s define a function that tells us if the given number is even or odd:

is_what(X) when X rem 2 == 0 -> even;
is_what(X) when X rem 2 /= 0 -> odd.

Notice the when <guard> part? We can use our new function like this:

is_what(3). % => odd
is_what(2). % => even

To achieve the same result in Ruby, we would do something like:

def is_what(x)
  return :even if x % 2  == 0

The cool thing is that guards can also be used in function heads, case and if statements.

Essentially they provide us with a succinct way to describe what we want out of a list.


Before we can do anything cool, we must really meet the most used data structure in Erlang, the List.

Lists are the bread & butter of many functional programming languages and are very efficient data types.

The first element of the list is called the head and the rest of the list is called the tail. Using pattern matching and the cons operator (|) we can easily extract those two out of a list:

> [Head|Tail] = [1,2,3,4,5].
> Head.
> Tail.
> [A|B] = Tail.
> A.
> B.
> ["Look what I did!"|B].
["Look what I did!",3,4,5]

To illustrate the recursive definition of lists, it helps to understand that all of the following notations:


are simply syntactic sugar for:


Note that the head of the list [1] is 1 and the tail is itself an empty list ([]).

Functional programming for real

Functions in Erlang are first-class citizens. This means we can manipulate them like any other data, passing them as arguments to other functions or making them the return values of other functions.

Let’s implement our own map/2 function. Turns out it’s very easy to do using lists and recursion:

map(F, [])    -> [];
map(F, [H|T]) -> [F(H) | map(F, T)].

Our map function accepts another function as it’s first argument and applies that function to every element in the list (ie. the second argument). It’s a high-order function.

The first definition is our base case which, when reached, the function will stop recursing and return the result. Applying a function to an empty list should return an empty list.

But what function should we pass to it? We can pass it a fun which will double the passed argument by 2. You can think of funs as anonymous functions, something like blocks in Ruby (well, not quite):

> map(fun(X) -> X*2 end, [1,2,3]).

That’s awesome! In two lines of code we created our own map function.

Let’s try passing our is_what/1 function that we defined earlier:

> F = fun(X) -> is_what(X) end.
> map(F, [1,2,3]).

Neat! Functions are first-class citizens in Erlang and we can manipulate them the same way we manipulate all other data. We can even define a function wich returns another function:

times(X) ->
  fun(Y) -> X*Y end.

We can now reuse our times/1 function:

> Double = times(2).
> Double(5).
> Triple = times(3).
> Triple(3).

Another common task is to filter out elements that meet a particular requirement, for example even numbers:

evens([]) ->
evens([H|T]) ->
  case H rem 2 == 0 of
    true -> [H | evens(T)];
    _    -> evens(T)

Let’s see it in action:

> hi:evens([1,2,3,4,5]).

Note that the order of definitions is important: each definition is evaluated from top to bottom, so if we wanted to add a clause to report back whatever invalid shape was received, we might do:

area(Other)            -> {unknown, Other};
area({circle, Radius}) -> 3.14 * Radius * Radius;
area({square, Side})   -> Side * Side.

But this would not work as expected, since the first clause would always match whatever argument we passed to area/1. The following demonstrates:

> area({shapeFromSpace, 5}).
> area({circle, 5}).
> area({rectangle, 5}).

To fix this we have to move the first clause to the bottom, in order to match only unknown shapes:

area({circle, Radius}) -> 3.14 * Radius * Radius;
area({square, Side})   -> Side * Side;
area(Other)            -> {unknown, Other}.

List comprehensions

It’s a very common pattern to map and filter lists. In other words to apply a function to every element of a list and from the resulting list, select those elements that meet a given requirement. List comprehensions provides us with a notation to achieve this result in a succinct yet powerful way.

Let’s say we want to pick all even numbers in a list and multiply them by 2. List comprehensions to the rescue:

> [X*2 || X <- [1,2,3,4,5], X rem 2 == 0].

Lists comprehensions are essentially tools to build and modify lists and are based on the set-builder notation.

That’s a fairly simple example. Let’s try something slightly more advanced by using multiple generators (ie. the part after the ||):

> [{X,Y} || X <- lists:seq(1,4), X rem 2 == 0, Y <- lists:seq(X,4)].

Here the expression is the tuple {X, Y}, then follows the generator X <- lists:seq(1,4) which essentially means X comes from lists:seq(1,4) (lists:seq/2 is a built-in function that generates a sequence of integers), then follows the guard X rem 2 == 0 and finally another generator.

As we can see, list comprehensions result in more succinct and idiomatic code.

That’s not all folks

In this short journey, we ran into lists, pattern matching, guards, functions and list comprehensions. What we learned can be summarized as follows:

  • List is the most useful data structure in Erlang. Common operations on them are very efficient: extracting the head or tail of a list and inserting an element at the beginning of the list are constant time operations. The cons operator is the basic tool for manipulating lists.

  • Pattern matching is a fundamendal characteristic of the language and a very powerful concept that is everywhere in Erlang programs. It results in more readable code (vs if-else statements). Guards are a great tool that provide additional power to pattern matching.

  • Functions are first-class citizens and are treated like any other data type. You can pass them as arguments, assign them to variables and use them as the return values of other functions.

This was a minimal introduction to some of the basic concepts of Erlang. The next post will be about the concurrency features of the language.

Discuss this post on HackerNews.

email · github · twitter · linkedin · keybase
PGP Fingerprint: D00B 5178 A5C0 C097