I still remember the time I spiked my first application in a functional language—it was Autumn 2018, I had come to 8th Light to start my journey as a crafter, and I didn’t even know what really made a functional language functional. Today, nearly one year later, I’m pretty confident it’s my favourite paradigm. In this blog post my intention is to share this little part of my journey, and the different things I learnt during it. Hopefully it may be of use to people who choose to start a journey like mine.

It isn’t always love at first sight

The first time I developed something in a functional language, I built a bot for Telegram in Clojure. That was the birth of Londibot—a bot that would ping me every time the Victoria line had a disruption. My mentor at that time gave me freedom to pick the project, and I thought it would be more fun to develop something that I could afterward put to good use, in this case for my commute. The language, though, he picked.

Dear reader, I’m not sure if you’ve ever tried any LISP dialect, but if you, like me, come from an Object-Oriented world, the learning curve can get very steep. I remember trying to figure out how to loop over lists to change the elements, trying to filter dictionaries… It’s very different than what we’re used to. The first things I got acquainted with that really paid off were the classic filter, map, and reduce; then came recursion to achieve loops, and many others. It was actually by learning these tools and giving them enough time to sink in that I started to understand the real difference between imperative and declarative code, and how the functional paradigm usually tends to favour more that declarative style. I could just say that learning a functional language and its idioms truly changed my perspective of how to solve problems and, ultimately, of how to express those solutions.

To summarize my first impressions, both the amount of parentheses that Clojure brought to the table (or any LISP dialect for that sake), and the amount of new concepts I had to learn, made it very overwhelming. It took me nearly a month until I started to enjoy the language a little bit more. The good thing is that at the end, it does pay off.

Keeping functions pure is crucial

One of the first mistakes I made early on in the project was not respecting the purity of the functions. In most OO languages, it is standard practice to create classes that are then instantiated into objects, which we then modify on demand. Say for example that you were to create a car parking app. You would probably think of the car park like the following:

var carPark = new CarPark();
var porsche = new Car("Porsche");
carPark.open();
carPark.park(porsche);
carPark.unpark(porsche);
carPark.close();

The problem with the above is that all state is saved within our CarPark instance, and as we park or unpark cars, that instance is modified. That’s called a side effect. A side effect is any operation that happens within a function that is collateral. Imagine a sum function that, apart from summing two numbers and returning the result, also saved it to a database. The saving to the database is a side-effect.

I’m aware that I still haven’t explained the concept of purity, but I wanted to talk about side-effects first because it greatly helps illuminate purity. In a nutshell, a function is pure when, for a same input, it always has the same output and never has any side effects. Printing a value to the log or simply changing some mutable state within the application makes our function impure.

You might be wondering: why is purity such a big deal though?. Purity makes our functions little sandboxes. It is by making pure functions that we can feel safe that when they change something it will only affect the context within the actual function. And this brings us to immutability. As we make our functions pure, that means that they only work with the parameters we provided them with AND they don’t mutate them, but return a result based on a copy of those parameters.

Now, that may sound confusing, but let’s say that we rewrote the above code to look something like this:

import java.util.ArrayList;
import java.util.List;
public class CarPark {
    private String status;
    private List<String> cars;
    public CarPark() {
        status = "Closed";
        cars = new ArrayList<>();
    }
    public CarPark(String status, List<String> cars) {
        this.status = status;
        this.cars = cars;
    }
    public CarPark open() {
        return new CarPark("Open", this.cars);
    }
    public CarPark close() {
        return new CarPark("Closed", this.cars);
    }
    public CarPark park(String car) {
        List<String> newList = new ArrayList<>(List.copyOf(this.cars));
        newList.add(car);
        return new CarPark(this.status, newList);
    }
    public CarPark unpark(String car) {
        List<String> newList = new ArrayList<>(List.copyOf(this.cars));
        newList.remove(car);
        return new CarPark(this.status, newList);
    }
}

Which when used would look like:

CarPark park = new CarPark();
CarPark openPark = park.open();
CarPark parkWithPorsche = openPark.park("Porsche");
CarPark parkWithoutPorsche = openPark.unpark("Porsche");
CarPark finalPark = parkWithoutPorsche.close();

Now, of course that is not idiomatic Java, but if we look at it closely–every time a function is executed, it provides us with a brand new instance with the attributes we expect. The major advantage of this is that if, say for example, another process were to run one of those methods concurrently, it would not screw up the operations running in other concurrent threads–each thread would be working with their own instances. Those functions are pure.

To wrap up all that I’ve just mentioned above, pure functions are functions that don’t create side effects and only affect the context within the function itself, not outside. As we keep our code pure, it is easier to test and overall more predictable. Purity keeps the code simple, while impurity increases the complexity.

Now, coming back to functional languages, Elixir specifically solves a lot of those problems out of the box! By default, all data structures are immutable in Elixir, which means that whenever they are altered, the runtime will give back a complete brand new copy of the structure. This means that if we were to implement a CarPark module, it would be:

CarPark.new() # Creates a new struct
|> CarPark.open() # Creates a new struct with status = open
|> CarPark.park("porsche") # Creates a new struct with a porsche parked
|> CarPark.unpark("porsche") # ...
|> CarPark.close() # ...

The code for that would be:

defmodule CarPark do
  defstruct[:status, :cars]
  def new(), do: %CarPark{status: :closed, cars: []}
  def open(%CarPark{} = park), do: %CarPark{park | status: :open}
  def closed(%CarPark{} = park), do: %CarPark{park | status: :closed}
  def park(%CarPark{cars: cars} = park, car_to_add) do
    %CarPark{park | cars: [car_to_add | cars]}
  end
  def unpark(%CarPark{cars: cars} = park, car_to_remove) do
    less_cars = Enum.filter(cars, fn car -> car == car_to_remove end)
    %CarPark{park | cars: less_cars}
  end
end

Now, like I just mentioned, most functional languages, like Elixir, come with immutability out of the box, and that is already a fantastic asset to help us keep purity throughout our codebase. Nonetheless, that module would stop being pure the moment I dump something along the lines of Database.save(carpark) in the functions. And if we did that, things like testing would go from being trivial to being complicated and requiring a lot of mental effort to keep track of the state of things.

Make functions read only horizontal or vertical

Changing to a lighter topic, I also wanted to talk about function readability. Most functional languages include some way of piping code—the pipe operator |> in Elixir or the threading macros -> and ->> in Clojure. This is a huge tool for making our code very readable, as it allows us to develop whole pipelines of transformations to our data in a very clean way. Say for example:

load()
|> format_head()
|> format_body()
|> convert_to_pdf()
|> upload(:google-drive)

The above piece of code is a beautiful example of how we can pipe multiple functions together, creating a highly readable function.

On the other hand, if we start mixing our style…

{:ok, document} = load()
pdf =
    document
    |> format_head()
    |> format_body()
    |> convert_to_pdf()
upload(:google-drive, pdf)

The above is a simple example, but try to read both functions a couple of times and think about the amount of effort required to follow either. They’re doing the same thing, but the first requires much less cognitive effort.

A good rule of the thumb is to go with either horizontal functions or vertical functions, and try to avoid mixing them as much as possible. There are many ways of achieving this, starting by trying to keep the function heads consistent in the order of parameters. For example, always accepting the struct to work with as the first parameter allows us to pipe functions much easier. Take a look at the upload/2 function above. By changing the order of the parameters we can either pipe it or not. A good example of this is the Enum module in Elixir. Just notice that for the Enum module to be pipeable it always accepts the enum as the first parameter and always returns a list!

[1, 2, 3, 4]
|> Enum.reverse()
|> Enum.map(fn x -> x*x end)
|> Enum.filter(fn -> x > 5 end)
|> Enum.reduce(fn x, acc -> x + acc end) # returns 25

For me, realizing this was huge. I remember when I developed my first modules with Elixir, I didn’t care much about the order of the parameters. After all, it’s not something you worry about that much in C# or Java, so most of my functions tended to read horizontally with an occasional vertical bit. When I started thinking about it, my thought process changed, and with it the readability of my code.

Reducers, gotta love ‘em

Following up on the order of parameters, come reducers. At the end of the day a reducer is simply a function that accepts some data, applies some transformations to it, and returns the modified version of the data. You’ll probably be thinking that it’s what all functions do, but the highlight here is that reducers don’t return the specific piece of data transformed, but the whole data WITH the transformed bits.

Say for example the String.capitalize/1 function:

iex> String.capitalize("abcd")
"Abcd"

It takes a string, capitalizes it, and returns the new string. When the function is applied, it gives us a new string with the A capitalized, not just the capital A. A different reducer that uses structs but achieves a similar purpose could look like:

def capitalize(%SomeStruct{value: value} = s) do
  %SomeStruct{s | value: String.capitalize(s)}
end

Instead of returning simply the new value, we return the complete structure, like for example IO.inspect/1. Some different reducers operating on the same structure could be:

def hash(%SomeStruct{value: value} = s) do
  hashed = :crypto.hash(:sha, value) |> Base.encode16
  %SomeStruct{s | value: hashed}
end
def downcase(%SomeStruct{value: value} = s) do
  %SomeStruct{s | value: String.downcase(s)}
end

Now, if we combine all these together, the value really sticks out. They allow us to do things like:

iex> %SomeStruct{value: "javier"}
     |> capitalize()
     |> hash()
     |> downcase()
%SomeStruct{value: "7f3d0970ec0e336aa08a9e14d4d88e79131e0065"}

That, my friends, is the magic of reducers. By applying transformations to our data structures and always making sure we return that same data structure, we can endlessly pipe our reducers, thus achieving the vertical functions we spoke about earlier.

Having the core business logic of our application modeled with reducers, as long as it makes sense, also makes it very easy to test:

def assert_result(%SomeStruct{value: v} = s, expected) do
    assert v == expected
    s
end
test "complex hashing works!" do
    %SomeStruct{value: "javier"}
    |> capitalize()
    |> assert_result("Javier")
    |> hash()
    |> assert_result("7F3D0970EC0E336AA08A9E14D4D88E79131E0065")
    |> downcase()
    |> assert_result("7f3d0970ec0e336aa08a9e14d4d88e79131e0065")
end

In my opinion, that is simply a beautiful test. I wish all my tests read in that clean and simple way. Reducers empower us to write better, cleaner code. It’s key though that the order of the parameters are consistent (so we can pipe our functions), and that they are also pure (so all the side effects are isolated in the boundaries of our application).

Encapsulate data with functions

Another lesson I learnt while coding with Clojure, early in my functional journey, was how to encapsulate data. One of the core pillars of OOP is encapsulation—keep your data private and your behaviour public. That way if some day you decide to modify the underlying implementation of your behaviour, it will never affect the consumers of your API.

In FP sometimes it can be slightly more challenging, or maybe just different, especially when the language doesn’t expose the concept of complex typing. I found out that by encapsulating data with functions, I was able to make my applications much more robust—consumers would have a consistent API with which to interact. In Elixir we have structs that solve that problem for us; but in Clojure, we don’t. So instead of passing an anonymous map around and accessing its keys, we could do something like:

(defn new-job [userid cronexpression service]
  {:userid userid :cronexpression cronexpression :service service})
(defn get-id [job]
  (:id job))
(defn get-cron-expr [job]
  (:cronexpression job))
(defn get-user-id [job]
  (:userid job))

Thus allowing consumers to access the data like:

(defn some-function [job]
  (-> job
      (job/get-user-id)
      (do-something)
      (do-something-else)))

By encapsulating our data sometimes with structs (but when not possible, with functions), we’re telling our consumers how we want them to use our API—we’re telling them, “Don’t use that property, because tomorrow it might not be there!”.

Wrapping up

Coming to an end, all I can say is… stay curious, keep learning, question yourself. Functional purity, encapsulation, and readability are all solutions to problems that emerged only after the code just didn’t click. All of this is just a quick summary of the different things I have learnt after having screwed up on many occasions. Londibot started as a Clojure bot when I was just starting to learn functional programming, and it was thanks to poor design choices and bad decisions that I kept reading and searching for better solutions to the problems that arose in my codebase. That’s how all this came up.