This content originally appeared on Jack Franklin and was authored by Jack Franklin
If you follow me on Twitter, or have read this blog for a while, you'll probably know that I'm a big fan of Elm. It's a functional, strictly typed language that compiles to JavaScript and is a great alternative to JavaScript for building web applications.
That said, it's not the only contender in this space. Reason is also a very popular option that has gained a lot of traction recently. I've always been interested in trying it out, and Advent of Code, a series of coding challenges posted each day leading up to Christmas, gave me a great excuse.
If you're into Elm, you might also be interested to know that I've done two videos completing Advent of Code challenges in Elm that you can find on Youtube.
If you're eager to skip ahead into the code, you can find it all on GitHub. In the rest of this post I'll talk you through my approach to getting up and running with Reason, and my thoughts on the language after trying it. I am not a Reason expert, so if you spot any errors or things I've misunderstood, please let me know! Equally, there might be better ways of solving the task, so if you have any suggestions please get in touch.
The first part of this blog post talks through my approach and how I solved the problem, and then we end with a list of my good and bad parts of trying Reason.
Getting started
I followed the official Installation and getting started guide to get easily up and running. It involved installing the compiler, BuckleScript, which is what takes Reason and produces JavaScript.
That let me run:
bsb -init my-new-project -theme basic-reason
To get a basic project up and running! I also installed reason-vscode so that I had nice error highlighting and type hinting as I coded. I find this particularly useful when working with a new language/framework that I'm not super familiar with.
Writing tests
I didn't want to build a UI to solve the Advent of Code problem; so I did a bit
of googling to see if I could use Reason to write some unit tests, and solve the
problem in a TDD style. I managed to find
bs-jest, a library that adds bindings to
BuckleScript to the JS testing framework Jest. This lets us write Reason, but
have it compiled into JavaScript that we can then run with Jest as normal. So
we'll write a tests.re
file, have it compiled into tests.js
, and then run
jest tests.js
. Setting this up was just a case of following the instructions
in the README, and it worked perfectly.
The Advent of Code challenge
I was taking on Day Two, and for this exercise only completed Part One. I'll leave Part Two as an exercise for you!
The first part of the exercise needed me to take a string, such as bababc
, and
calculate the frequencies that letters occur. So for this string, we'd end up
with:
{ a: 2, b: 3, c: 1 }
So that was the first thing I set out to write. I discovered that BuckleScript
provides a
Js.Dict
module
that is the equivalent of a native JS object, and I could use that. It also
provides
Js.Array
, and
Js.String
.
Using a combination of methods from these modules, I could split my input, and
loop over it, updating a dict with new frequencies as I go through each letter.
I decided to store the frequencies in a dictionary. In Reason you have to decide what the types of the values are in a dictionary, so I went with integers, given we're counting frequencies.
I first set out to write a function that could take a dictionary and a letter, and update the frequency for that letter:
- If the letter has no entry in the dictionary, create one and set the frequency to one.
- If the letter has a frequency, update the count by one.
Defining this function looks very similar to JavaScript:
let incrementOrSetFrequency =
(frequencies: Js.Dict.t(int), letter: string): Js.Dict.t(int) => {
};
The bit that Reason adds is the type annotations. After each of the two arguments, we declare the types. We don't have to do this - Reason will be able to infer them for us - but I find it helps me work with code if I've documented the type, and very rarely the compiler can infer a type slightly differently to what you actually want it to be.
The type annotation above says that frequencies
is a Js.Dict.t(int)
, which
means a dictionary where each value is an int
type. letter
is a string
.
After the arguments we have the return type, which is also a dict, as we want to
take the dict, update it, and then return it again.
The first thing we need to do is check to see if letter
is in the dictionary,
and we can use Js.Dict.get(frequencies, letter)
to do this. It doesn't return
the value or undefined
though, like you would expect in JavaScript. Instead,
it returns something that's an Option
type. This is Reason's way of trying to
avoid unexpected undefined
or null
s in your application. You can read more
about
Option
on the Reason docs.
When you have a function that returns an Option
type, you can use
pattern matching to see
what the value is, and act accordingly. So if we look in our dictionary for our
letter and it returns None
, we need to add the letter. If it returns
Some(int)
, we want to increment it by one:
let incrementOrSetFrequency =
(frequencies: Js.Dict.t(int), letter: string): Js.Dict.t(int) => {
switch (Js.Dict.get(frequencies, letter)) {
| Some(x) =>
Js.Dict.set(frequencies, letter, x + 1);
frequencies;
| None =>
Js.Dict.set(frequencies, letter, 1);
frequencies;
};
};
Getting our first test passing
At this point I decided I'd figured out enough Reason to be dangerous, and
wanted to write a test so I could work towards getting it passing. I created
__tests__/daytwo_test.re
:
open Jest;
describe("DayTwo", () => {
open Expect;
test("letterFrequencies", () =>
expect(DayTwo.letterFrequencies("bababc"))
|> toEqual(Js.Dict.fromList([("b", 3), ("a", 2), ("c", 1)]))
);
If you've written JS tests with Jest, you'll probably find the above quite
intuitive, and I was able to use Js.Dict.fromList
to take a list of tuples and
create the dictionary that I needed for the test. The compiler compiled this
into a JS file that I could run using the regular Jest CLI. This was one thing I
liked about Reason; I can use the regular Jest CLI, rather than having to use a
special one specifically for Reason. Jest's CLI is so good that it makes total
sense to work on top of it rather than creating a language specific one from
scratch.
To get the test passing we needed to take our input string, split it into a list
of letters, and run each one through our incrementOrSetFrequency
function:
let letterFrequencies = (input: string): Js.Dict.t(int) => {
let frequencies = Js.Dict.empty();
input
|> Js.String.split("")
|> Js.Array.reduce(
(acc, currentValue) => incrementOrSetFrequency(acc, currentValue),
frequencies,
);
};
And with that the test is passing!
Getting frequencies for our entire puzzle input
Next we need to take our full puzzle input, which is a series of strings, and run the above function on each of them, so we can start to work towards the final answer that we need.
Once again, I start by writing a test. I replicate the input that the real puzzle provides by putting each entry on its own line. I want to make sure we get the logic for splitting lines works properly.
Note that {|string here|}
allows us to define a multi-line string.
test("checksum", () => {
let puzzleInput = {|
abcdef
bababc
abbcde
abcccd
aabcdd
abcdee
ababab
|};
expect(DayTwo.checksum(puzzleInput)) |> toEqual(12);
});
We can use the familiar Js.String.split
once again here, but pass it "\n"
as
the thing to split on. We then map the resulting lines over String.trim
, which
trims any whitespace and removes it. Note that we're not using
Js.String.trim
here, this is the
ReasonML module String
, not
the
BuckleScript Js.String
module.
This was one of the things I found most confusing when learning Reason. It
wasn't clear why some of the functions we use are Reason modules, and others are
provided by BuckleScript.
If you're familiar with Reason and can clarify the above confusion, I'd love to talk it through and update the blog post to include it.
So, the first part of the checksum
function is to take the multi-line input,
split it, and then ensure that we don't have any blanks:
let checksum = (input: string): int => {
input
|> Js.String.split("\n")
|> Js.Array.map(String.trim)
|> Js.Array.filter(s => String.length(s) > 0)
// note: this is invalid (we're not returning an int)
Once I've split the lines and given them a trim, I then use Js.Array.filter
to
remove any strings that are entirely empty. Now we are working with an array of
letter frequencies that looks something like this:
[
"abcdef",
"bababc",
"abbcde",
"abcccd",
"aabcdd",
"abcdee",
"ababab",
]
So we want to take each one and pass it into the letterFrequencies
function
that we have defined:
let checksum = (input: string): int => {
input
|> Js.String.split("\n")
|> Js.Array.map(String.trim)
|> Js.Array.filter(s => String.length(s) > 0)
|> Js.Array.map(letterFrequencies)
// note: this is invalid (we're not returning an int)
Now we've turned that list of strings into a list of frequencies. This code sample highlights one of my favourite Reason features (I'm biased as it's also a favourite feature of mine from other functional languages like Elm and Elixir), the pipeline operator. The pipeline operator takes the thing on the left and passes it as the last argument to the function on the right. It means fewer parentheses around everything and lends itself to creating really readable code.
Calculating frequency occurrences
Now we have a list of frequency dictionaries, we need to take them and figure out:
- how many of them contain a letter exactly 3 times
- how many of them contain a letter exactly 2 times
The result for each of those is what we'll need to multiply together to get our checksum, which is the solution to our puzzle.
What I'd like to do is take our list of frequencies and map it into a list of
Reason objects that contain two properties, twice
and thrice
. These will be
booleans and correspond to if a word contains a letter twice or thrice. To help
the compiler give me good type errors if I make a mistake, I create a custom
type:
type twiceAndThriceFrequency = {
twice: bool,
thrice: bool,
};
This declares a type, twiceAndThriceFrequency
, which is an object with two
properties that are both booleans. I can then create a function that will take a
frequencies dictionary and convert it into one of these objects. Now I have this
custom type, I can use it in the type annotation too:
let findTwicesAndThrices = (frequencies: Js.Dict.t(int)): twiceAndThriceFrequency => {
{twice: true, thrice: true }
};
For now I've hardcoded the values to both be true
, we will fill those in
shortly. Notice how having the custom type defined makes the type annotation
read really nicely and clearly.
To figure out the value of the twice
and thrice
keys, we need to see if the
frequencies dictionary has any values of 2
or 3
in it. For this problem, we
don't actually care about which letter occurs two or three times, we just need
to know if any of them do.
We can use Js.Dict.values
, which takes a dictionary and returns an array of
the values inside it. It's just like Object.values()
in JavaScript. We can
then use Js.Array.some
, which takes an array and a function and tells us if
any items in the array satisfy it. Therefore, we can define the functions
hasTwices
and hasThrices
like so:
let hasTwices = (frequencies: Js.Dict.t(int)): bool => {
frequencies |> Js.Dict.values |> Js.Array.some(v => v === 2);
};
let hasThrices = (frequencies: Js.Dict.t(int)): bool => {
frequencies |> Js.Dict.values |> Js.Array.some(v => v === 3);
};
Note that in this solution I'm not worrying about performance. If I was, we'd be doing this differently to reduce the number of times we iterate over the
frequencies
array. I'll leave it as an exercise to the reader to improve that.
Mapping to our twiceAndThriceFrequency
type
Now we have these functions, we can define a function that will take a
frequencies dictionary and return a twiceAndThriceFrequency
type:
let findTwicesAndThrices = (frequencies: Js.Dict.t(int)): twiceAndThriceFrequency => {
{twice: hasTwices(frequencies), thrice: hasThrices(frequencies)};
};
Notice that we don't need the
return
keyword in Reason. The last expression in a function is automatically returned for you.
And once we have this function, we can update our main checksum
function:
let checksum = (input: string): int => {
input
|> Js.String.split("\n")
|> Js.Array.map(String.trim)
|> Js.Array.filter(s => String.length(s) > 0)
|> Js.Array.map(letterFrequencies)
|> Js.Array.map(findTwicesAndThrices)
// note: this is invalid (we're not returning an int)
Calculating our checksum
At this point we are working with a list of objects that have
{ twice: true/false, thrice: true/false }
within them. We want to go through
this list and reduce it down to two values: the number of times that we have a
letter occurring twice, and the number of times we have a letter occurring three
times. So if we have this list:
[
{ twice: true, thrice: false },
{ twice: false, thrice: false },
{ twice: true, thrice: true },
]
We want to end up with:
{ twice: 2, thrice: 1 }
It's then these two numbers that we multiply to find our checksum.
We can use Js.Array.reduce
to do this. It will take our array and loop through
each value in turn, allowing us to check the values of twice
and thrice
and
increment our accumulator accordingly. Our starting accumulator will be an
object, which I also define a type for:
type twiceAndThriceCounter = {
twice: int,
thrice: int,
};
And now we can start planning our reduce
call:
|> Js.Array.reduce(
(acc: twiceAndThriceCounter, currentValue: twiceAndThriceFrequency) => acc
{twice: 0, thrice: 0},
)
Inside the body of the callback function, we need to check the currentValue
and check the values of twice
and thrice
.
This is a case where Reason's pattern matching comes in really handy. We can write code that pattern matches against the object and its values:
switch (currentValue) {
| {twice: true, thrice: true} => {
twice: acc.twice + 1,
thrice: acc.thrice + 1,
}
| {twice: true, thrice: false} => {
twice: acc.twice + 1,
thrice: acc.thrice,
}
| {twice: false, thrice: true} => {
twice: acc.twice,
thrice: acc.thrice + 1,
}
| {twice: false, thrice: false} => acc
},
Each case that we're matching against starts with the pipe (|
) and then we
match against the twice
and thrice
values within currentValue
. So the
first will match only if currentValue
has both values set to true, in which
case we increment both of our counters. In the case of one of twice
or
thrice
being true, we increment the appropriate counter and if both values are
false
, we do nothing.
Pattern matching is my favourite feature of Reason (it's also one of my
favourite parts of Elm), and it leads to some really nice, expressive code.
What's also nice is that if we don't write code that deals with every possible
case, we get a compiler error. In the example below, I've removed the case that
deals with both values being true
. You can see the compiler spot this and tell
me:
Warning number 8
/Users/jackfranklin/git/advent-of-code/day-two-reason-ml/src/DayTwo.re 55:10-65:10
53 ┆ |> Js.Array.reduce(
54 ┆ (acc: twiceAndThriceCounter, currentValue: twiceAndThriceFrequenc
y) =>
55 ┆ switch (currentValue) {
56 ┆ | {twice: true, thrice: false} => {
. ┆ ...
64 ┆ | {twice: false, thrice: false} => acc
65 ┆ },
66 ┆ {twice: 0, thrice: 0},
67 ┆ )
You forgot to handle a possible value here, for example:
{twice=true; thrice=true}
This means you can never end up with code in production that doesn't deal with all possible cases, which is fantastic. It also means if you refactor and now your pattern matching is out of date, the compiler will tell you.
Once we have this reduce done, it's going to end up turning our array of frequencies into one object with two values. The solution to the puzzle (and what we need to get our test passing) is to take these values and multiply them. We can do this by piping our object into an anonymous function that does just this:
|> result => result.twice * result.thrice
And with this, our tests are back to green!
PASS __tests__/daytwo_test.bs.js
DayTwo
✓ letterFrequencies (6ms)
✓ checksum (1ms)
There's one small refactor we can make here though. Much like JavaScript and its ES2015 destructuring, we can destructure an object into the keys when it's passed into a function. So we can rewrite our final line as:
|> (({twice, thrice}) => twice * thrice)
Which I think reads a bit more clearly. And with that, our puzzle is solved!
Conclusion
This was literally the first time I'd written Reason and after finishing the Advent of Code challenge I took a moment to think through what I found good, and what I struggled with, from the perspective of a beginner using a new language.
It's also worth noting that my experience with Elm almost certainly makes it easier for me to learn Reason, there are similarities between the two.
Things I liked
- The tight interopability between Reason and JavaScript is very compelling. I could easily see myself writing one module in Reason in an existing JS application because the interop is so smooth and easy.
- Continuing from the previous point, the fact that Reason can use Jest for its test runner is excellent. Not having to learn how to run another test runner was a major bonus. It also helps that Jest is absolutely exceptional and packs in a tonne of useful features, so it makes perfect sense that Reason would lean on that rather than build out a brand new test runner.
- On the whole I found compiler errors clear and obvious. One of my main gripes with TypeScript is that some of the compiler messages were hard to parse, but Reason gave me understandable messages that I really appreciated, particularly as a beginner.
- The documentation on the Reason site is excellent. Take this page on pattern matching as an example: it's clear, the code samples are easy to follow, and it explains things thoroughly. It also avoids any complex jargon and doesn't attempt to sound super clever.
- This one is editor specific, but the reason-vscode plugin gives a really good developer experience. It was easy to quickly get formatting, syntax highlighting, compiler errors and so on in my editor. (If you use another editor, there are links to plugins on the Reason site).
- Reason includes
refmt
, a code formatter for Reason code. Much like Prettier for JavaScript, this runs and formats your code. What's great about this is that all Reason projects use this, so all Reason code is formatted the same, and that as a beginner any worries about conventions or how to format something are gone. I just run the formatter! The VSCode plugin runs this for me when I save, so I just didn't have to think about it.
Things I found confusing
Please remember that I am writing this as a Reason beginner, not an authority! If I've misunderstood something or made a mistake, please let me know and I'd be happy to update the blog post and give credit accordingly.
- I've struggled in my head to fully understand the iteraction between Reason, OCaml and BuckleScript. In my head Reason is a syntax on top of OCaml, and BuckleScript is the compiler that can produce JavaScript. I'm not sure if my mental model stacks up though, and I found it hard to get clarity on this online. Update!: Axel was kind enough to share this diagram which I think makes things clearer and provides a nice picture.
- I also found it confusing where to look for documentation for available modules. For example, when wanting to split a string, I found the Str Reason module. However, this isn't available when compiling with BuckleScript, so I ended up using the docs from the BuckleScript API for Js.String. After this I was confused as to which one I should use, and why some modules exist in BuckleScript, but others in Reason. This is still a big point of confusion for me - if you can help me understand it I'd love to chat and also update this blog post!
- I think this is me being strongly biased based on my Elm experience, but I
didn't love that methods like
Array.get may raise an exception
if the item at the given index isn't present. I think here I'm projecting my
expectations from Elm onto Reason, and actually the approach Reason has taken
probably is an easier entry point for JS programmers, but I'd rather they all
return the
Option
type, which Reason does support and use
All in all, I'd really recommend giving Reason a go! I'm excited to see where the language and ecosystem goes in 2019 and beyond, and I'll definitely be playing with it some more, maybe next time on an actual frontend project, rather than just a coding exercise.
This content originally appeared on Jack Franklin and was authored by Jack Franklin
Jack Franklin | Sciencx (2019-01-23T00:00:00+00:00) Adventures with ReasonML. Retrieved from https://www.scien.cx/2019/01/23/adventures-with-reasonml/
Please log in to upload a file.
There are no updates yet.
Click the Upload button above to add an update.