Ideas and guidelines for architecting larger applications in Elm to be modular and extensible

Architecture.md

Architecture in Elm

This document is a collection of concepts and strategies to make large Elm projects modular and extensible.

We will start by thinking about the structure of signals in our program. Broadly speaking, your application state should live in one big foldp. You will probably merge a bunch of input signals into a single stream of updates. This sounds a bit crazy at first, but it is in the same ballpark as Om or Facebook's Flux. There are a couple major benefits to having a centralized home for your application state:

1. There is a single source of truth. Traditional approaches force you to write a decent amount of custom and error prone code to synchronize state between many different stateful components. (The state of this widget needs to be synced with the application state, which needs to be synced with some other widget, etc.) By placing all of your state in one location, you eliminate an entire class of bugs in which two components get into inconsistent states. We also think you will end up writing much less code. That has been our observation in Elm so far.

2. Save and Undo become quite easy. Many applications would benefit from the ability to save all application state and send it off to the server so it can be reloaded at some later date. This is extremely difficult when your application state is spread all over the place and potentially tied to objects that cannot be serialized. With a central store, this becomes very simple. Many applications would also benefit from the ability to easily undo user's actions. For example, a painting app is better with Undo. Since everything is immutable in Elm, this is also very easy. Saving past states is trivial, and you will automatically get pretty good sharing guarantees to keep the size of the snapshots down.

I think these two strengths will be extremely worthwhile in large applications, though I feel that strength 1 is a huge deal for speeding up development and avoiding silly bugs that waste your time.

So most of your code will be pure functions that make your big foldp actually do the right thing. The rest of this document focuses on how to make that code modular and extensible.

Modularity

To make things modular, the major strategy is to hide implementation details, as shown in this pseudocode. When you create a widget, this makes it possible to expose exactly the API you want. Essentially just this kind of info:

type Model.State
type Model.Action

Model.initialize : String -> ... -> State

Update.step : Action -> State -> State
Update.resetField : State -> State

View.full : Input Action -> State -> Element
View.mini : Input Action -> State -> Element

As a user, you don't know anything about what State really is, but you have carefully selected functions for creating it, stepping it, and doing custom modifications without an Action (e.g. resetField) in case other components need to act on the state. The designer has full control over the API they expose and can hide any details they want.

In Elm you have the added benefit that these abstraction boundaries are quite strong. Unlike in JS, you cannot just inspect the structure of arbitrary values and do what you want with that. That means best practices must be enforced with culture or dogma. In Elm you can actually ensure that people write code in a good way.

Extensibility

When you have a ton of widgets, all with different sets of actions, how do you use them all together? The TodoMVC code uses a big ADT called Action, but that is not easy to extend. There are a few useful techniques here.

Nesting

So let's say we have three different widgets, each with their code living in modules called SearchBar, Filters, and Results. This means we have three sets of actions SearchBar.Action, Filters.Action, and Results.Action which we do not know anything about. To put them together, we would create some nested actions and state:

data Action
    = SearchBar SearchBar.Action
    | Filters Filters.Action
    | Results Results.Action

type State =
    { searchBar : SearchBar.State
    , filters   : Filters.State
    , results   : Results.State
    }

step : Action -> State -> State
step action state =
  case action of
    SearchBar a -> { state | searchBar <- SearchBar.step a state.searchBar }
    Filters   a -> { state | filters   <- Filters.step   a state.filters   }
    Results   a -> { state | results   <- Results.step   a state.results   }

You can just keep nesting and nesting like this. For example, the Results module may be made up of 4 smaller widgets with their own actions and state and we don't need to know anything about those details.

Generalizing Actions

In some cases where you want an ADT to be more extensible. Here is the fully general approach:

type Action = State -> State

removeTask : Int -> Action
removeTask id state =
    { state | tasks <- filter (\task -> task.id /= id) state.tasks }

We can then expose a set of functions like removeTask that let people create a certain set of Actions. So rather than having a step function that handles the many cases of an ADT, we just apply the Action to the state:

step : Action -> State -> State
step action state = action state

It is just function application! This means people can write their own functions like (resize : Int -> Int -> State -> State) just by building up from the publicly available ways to transition state. This means we can act on State in composable ways without revealing any facts about State.

This can also be mixed with the ADT approach like so:

data Action
    = RemoveTask Int
    | ...
    | Anything (State -> State)

This way you still have the structure of an ADT that helps organize code and forces you to think hard about what functionality you actually want, but you can escape that structure if needed using the Anything case which will just step the state in an arbitrary way.

I am not sure how this approach will work out in practice as I personally favor ADTs, but I think we should explore it more and see how it goes. I can see it going either way.

Reusable Actions

So assuming we go with something like type Action = State -> State, how can we reuse an action on different kinds of state? Elm's extensible records can help a lot with this. Rather than defining a totally opaque type State, we can build it from smaller pieces:

type Position r = { r |  x:Float,  y:Float }
type Velocity r = { r | vx:Float, vy:Float }
type Lives    r = { r | lives:Int }
type Coins    r = { r | coins:Int }

type Mario  = Position (Velocity (Lives {}))
type Goomba = Position (Velocity {})
type Brick  = Position (Coins {})

oneUp : Lives r -> Lives r

gravity : Time -> Position (Velocity r) -> Position (Velocity r)

takeCoin : Coins r -> Coins r

Now we can use the oneUp and gravity functions on anything that have the required fields. So stepping Mario's state forward could look like oneUp . gravity dt. This gives us an interesting way to reuse functions that may be more clever than is necessary in practice, but at least now you know it's available to you!

Specializing Inputs

This section is about conveniences that may need to be added to Elm to fully realize the vision outlined here.

I think it may be pleasant/necessary to introduce a function something like this:

specialize : (particular -> general) -> Input general -> Input particular

Notice that the order of arguments seems a bit wonky here. You are creating a new input of particulars and you need a way to convert all of those to the more general type to integrate with the rest of the system. The idea is that you can create one input, but have all 3 of your widgets report to it with things like this:

-- using the general Action ADT defined earlier
searchBarInput = specialize SearchBar appInput
filtersInput   = specialize Filters   appInput
resultsInput   = specialize Results   appInput

It may also be a good idea to make input creation syntactic as with ports and to demand that they all be created in the Main module to ensure that people structure their code in a reusable way.

Focus

It is possible to give widgets access to a small part of the overall state. Say you want the Filters widget to set some state relevant to the SearchBox widget. One way to do this is to have a Filters.Action that knows about that particular piece of state. You could handle that particular action one level above in the code that manages the many subcomponents. I think this is fairly intuitive and will work well enough.

A Focus generalizes this idea of giving access to a subpart of a big data structure. I have tried to outline the idea as I think it should look in Elm, but I'm not sure it is actually a big win for architecting applications. It sounds promising, but it has a conceptual complexity cost that I'm not sure will pay off. My main concern is that, by making it easy to look deep inside of data structures, it encourages you to stop thinking about how to make these substructures modular, perhaps leading to an architecture that is not as nice and has extra conceptual complexity.

The reason for my skepticism is that I'd like to keep the number of new concepts as low as possible. If we can get by without this, I think Elm will be more attractive.