Does Elm Harmonize With F#?

Diving Into Another ML Language

After experimenting with Elm, I was hooked. I could feel that working with a smart type system brought both reliability, in the sense of "it does what I want it to do", and safety, in the sense of "it does not have bugs".

Elm's main focus is the front end (even though some work is being done), so a natural next question could be What would using an Elm-like language in the server be like? And by "Elm-like" I mean a language with immutable data structures as default, encouraging expressions over statements, and of course, having those cool discriminated unions we all learned to love. :)

After doing some quick research, my finalists were Haskell, Ocaml and F#. A former coworker was always saying good things about F#, and after seeing things like FSharp.Data, Xamarin and this amazing blog, I decided to go with F#.

Setting Up

The first week with F# was not easy. I was a little clueless about where to get started since I don't have a Windows machine. So, I downloaded Xamarin Studio, and used it for a couple of days. It's a good IDE with a so so editor, so I started looking for some alternatives.

Every F# project needs a *.fsproj config file, which is a big messy XML that is not meant to be changed by "human beings" (as opposed to the simpler package.json from Node or elm-package.json from Elm). But it turns out that the F# open source community has built some nice tools to deal with it. Summing up, install the Ionide extension to VS Code or Atom, and everything will be easier.

My programming workflow then became: start coding something, and "send" the code little by little to the REPL. Define some types, send to the REPL. Write a function, send to the REPL. Write some use cases, send to the REPL. It's the exact same workflow I used with Clojure, and it leads to a lot of early feedback and agility from the very beginning.

With F# up and running, I decided to implement a simple project to learn the language.

The Project

My wife was a girl scout when she was a kid. She told me about a common girl scout activity in the US, which is selling cookies door to door to raise money for their girl scout activities. I found that was a cool tradition, so I decided to implement a girl scout cookie selling dashboard! :)

The requirements are: each scout somehow send commands like "Visit House" or "Sell 5 Cookies" to the server. The server validates the command, and broadcast events like "Maggie Just Visited A House" or "Lisa Just Sold 5 Cookies" to the dashboards, that are updated in realtime and consolidate all the information.

The Domain Model

Each scout can be modelled as a "state machine". This is where these type systems really shine. F#, like Elm, has union types and pattern matching, so the code was pretty similar. For instance:

type ScoutState =
    | Walking
    | Visiting
    | HavingFun

type ScoutCommand =
    | StartDay
    | VisitHouse
    | Sell of int
    | HaveFun

type ScoutEvent =
    | DayStarted
    | HouseVisited
    | Sold of int
    | DayFinished

type ScoutError =
    | ShouldBeHavingFun
    | ShouldBeWalking
    | ShouldBeVisiting

A word about Commands and Events: when programming in Elm, "everything" that happened in the application was called a Message. The Update function understands the Messages, updates the state accordingly, and then emits or not new Messages.

When reading about F#, I came across lots of reading materials on Domain Driven Design (DDD), Event Sourcing and Command Query Responsibility Segregation (CQRS). One of the nice ideas I've read is the separation between Commands and Events among the Messages.

Every time a user wants to change the domain, it issues a Command. If the Command actually changes the domain, an Event is generated. An Event represents things that actually happened to the domain. For example, if a scout issues a Command Sell 3, the server checks if it's a valid Command for the actual State, and if it's ok it issues an Event Sold 3. If it's not valid, let's say the scout was not visiting any house, it returns a ShouldBeVisiting error, and the State is not updated.

The cool part is that we can store all the Events, and query them at will. The present State is a replay of all the past Events! That's what Event Sourcing is all about, and I really recommend watching every video you can from Greg Young, the "father" of the Event Sourcing pattern.

With that in mind, this is the final state machine representation of the scouts:

Scout's Model

The final code for the domain is here.

A Simple Event Store

The events produced by the server need to be stored somewhere. To continue the learning experience, I decided to implement the simplest in-memory event store I could think of. It should be able to store every event, broadcast them to listeners, and that's it.

Since the store actually stores data, and this data is accessed through methods, I thought that modelling the store as an object would be ok. Yes, it's the kind of object that we try to escape with functional programming, but it still has it uses :)

It was also good to see that "object oriented programming" with F# is very simple, and has almost no boilerplate. You just declare a type with () near the name, and declare "members" of the type. For instance:

type EventStore() =
    let eventList =
        new ResizeArray<String * ScoutEvent>()

    member this.Save(name, events) =
        events |> List.iter (fun e -> eventList.Add(name, e))

    member this.Get() =
        eventList

When instanced with let store = new EventStore(), this object creates an array of String * ScoutEvent tuples. This will hold all the events produced by the application, together with the scout's name that produced the event.

F# has a very interesting feature called computation expressions. The docs describe it as a convenient syntax for writing computations that can be sequenced and combined using control flow constructs and bindings. I do not know of an easier way to describe it, so let me show two examples, async and seq:

// First example

// this works just like Javascript's async/await,
// or Clojure's core.async
let asyncFetchedDocument = async {
    let url = "http://some-url.com/api"

    // this line will wait for the response without blocking the thread
    let! res = someAsyncFetchFunction url

    return res.data }

let fetchedDocument =
    Async.RunSynchronously asyncFetchedDocument


// Second example

// this produces a lazy sequence
let lazySeq = seq {
    for i in 1 .. 5 do
      if (i % 2 = 0) then
        yield i
        yield -i }

let list =
    Seq.toList lazySeq
// => [2; -2; 4; -4]

In the event store I wrote, I used a query computation expression to get all the events from a given name in the EventStore:

...

member this.Get(name) =
    query {
        for (n, ev) in eventList do
        where (n = name)
        select ev
    } |> Seq.toList

...

That's really cool, right? It feels like a strongly typed SQL :) I feel that Elm would strongly benefit from something like this.

So, to finish the EventStore, I needed a pub/sub to the saved events:

...

  let saveEvent =
      new Event<'Key * 'Event>()

  member this.SaveEvent =
      saveEvent.Publish

  member this.Save(name, events) =
      events |> List.iter (fun e -> eventList.Add(name, e))
      events |> List.iter (fun e -> saveEvent.Trigger((name, e)))

...

The events are published at SaveEvent, and saving an event triggers the listeners. An example subscription would be:

let logger ev = printfn "-- EVENT: %A" ev

store.SaveEvent.Add(logger)

The complete EventStore, with generic typing, is here.

A Server

Years ago I worked with C#, and wrote a server with ASP.NET MVC. In this project, I thought I was going to write the server using F#/C# "interop" (and I was not very happy about it :) ). Fortunately, I came across an "F# native" web server framework called Suave. Suave in Portuguese means smooth, and that's exactly how it feels to write a server with this framework!

For instance, look at the final code for the server:

let app store =
    choose [
        path "/"          >=> Files.sendFile "build/index.html" true
        path "/dashboard" >=> Files.sendFile "build/dashboard.html" true
        path "/websocket" >=> handShake (eventSocket store)
        path "/command"   >=> POST >=> request(commandRequest store)
        ServerErrors.INTERNAL_ERROR "Sorry, route not valid!" ]

Really clear, right? Writing this server was very easy. The routes are defined through function composition, and Suave comes with some functions that define standard server behaviors, like Filters.POST, Files.sendFile and ServerErrors.INTERNAL_ERROR.

I found one problem while trying to make it work: there's simply no documentation on working with websockets. After googling it, the material I found was from a couple of StackOverflow questions, and this example from the project's github.

It took some time, but, after a few hours, I was able to make it work. This is part of the final code for the socket code:

...

socket {
    let loop = ref true
    while !loop do
        let! msg = webSocket.read()
        match msg with
        | (Ping, _, _) -> do! webSocket.send Pong [||] true
        | (Close, _, _) ->
            do! webSocket.send Close [||] true
            subscription.Dispose()
            loop := false
        | _ -> ()
    }

It's a computation expression! Gotta love those.

The complete code for the server can be found here.

An Elm dashboard

After implementing F#, it was time to return to Elm. The dashboard architecture was simple: the app listens for events through the native websockets implementation, and it drives the standard Elm architecture. This is what the final result looks like (notice I'm using Postman to send the scout's commands to the server):

Event Source A Event Source B

One of the best parts of Elm is that the structure of the code does not seem to change from application to application. The "framework" is so close to the language itself, that one simply does not try to do things differently - I'm not even sure it's possible. Because of that, we can focus more on the real complexity of the problem, and think less and less about implementation details.

The only different part from my earlier learning Elm series is the websockets part. But it's really, really simple:

type Msg
    = EventReceived String

subscriptions : Model -> Sub Msg
subscriptions model =
    WebSocket.listen scoutEventsServer EventReceived

And now I only have to treat EventReceived messages in the update function.

Something I did not like: I felt like I had to repeat a lot of the code in the client and the server, mostly in the models. Mismatches between client and server models are a huge source of errors. Although I don't think it would happen here - after all, they are in the same project :) - I would have to update two pieces of code each time the model needed to change.

The complete code for the dashboard is here.

Final Conclusions

In the end, I really enjoyed the experiment. Even though the languages appear similar, I could see a lot of differences between the two, mostly due to Elm being a language looking for reliability in a narrow scope, and F# looking for pragmatism in a much larger scope.

Ways Elm Is Superior To F

Ways F# Is Superior To Elm

Things I Learned In The Process

I learned some valuable lessons working on this project. First of all, on the language level, I learned a lot about how to use types better. ML-style languages make it easier to use types to actively make your code more reliable. I recommend reading this blog post series as a masterclass in "making illegal states unrepresentable". It's a very powerful and simple idea that consists of using the type system as way to maintain the application in a valid state 100% of the time. Another way of phrasing it is "testing is good, but impossible is better" :)

And second, in the architecture level, I learned about Event Sourcing and CQRS. Today, I really believe that using event logs as first class citizens is the way to go in most situations. Even in this small project, I could feel the benefits: whenever I was coming up with a new way of visualizing the scouts information in the dashboard, I never had to change any schema, or any code in the backend. "Read models" are only different queries in the stored events.

I feel that in a lot of companies, our main storage consists of "last states of a bunch of view models", and we spend a lot of money on different analytics, and log-managing tools to deal with the really important business questions. If these companies already stored events, and treated them as they treat their main data, an immense value would be generated at a much lower cost.

Next Steps

The repeated code between client and server made me curious about trying something more "unified", so I'll try some Fable for frontend programming.

Also, I'll also try to use more Event Sourcing whenever I can :)

September 8, 2016.