30.09.2020

This article will discuss app state and show one approach to it in Rust, continuing the TicTacToe application example and using the valerie framework. This is the third in a series which starts here and continues from where we left off last time.

The code for this example is here: my fork of valerie/examples/tictactoe. (Actually this commit is a little ahead of this article)

Why you might want to lift state out of components§ 

Fulcro made this clear to me.

When one object appears at multiple locations in a page, the information needs to be stored at a top-level component and passed down (as props in React and clones of the State structs in valerie). This couples components and their nesting, to thread the state object from where it is stored, to where it is displayed.

When a deep sub-component references a piece of state, this must be known to each wrapping component.

For example, a nested view of order lines needs to make a decision about how to present information about the product it relates to. The order line state may know just the foreign key identifier of the product the line relates to, but the user may be expecting its short code, display name and packaging dimensions. It would be better if the decision about what to display could be delegated to the order line view, without asking order to pass down too much (any?) product information (and then, what if we want to join to stock on hand?).

Returning to our implementation of Tic Tac Toe, if our game was constructed like this:

fn game() -> Node {
    div!(
        div!(
            board(),  // NOTE: three boards!
            board(), 
            board(),
        ).class("game-board"),
        div!(
            div!(),
            ol!()
        ).class("game-info")
    ).class("game").into()
}

we would have had three independent games proceeding in parallel.

A different intention would be for each of those boards to reference the same playing state - various views of the same data, remaining synchronised. Various view functions could show different perspectives on the running game. We could answer questions and put that information anywhere on the page, for example: how many turns done? Win possible next turn? How many doubles?

(Later we will explore how to select which intention is rendered.)

If we have our state completely separate from views, how would it look?

Static-scope state we might call "Store"§ 

In Rust, static-scoped items are accessible from any part of the application without being passed as parameters. What if the rest of state was accessible through a lookup to static stores, using an identifier?

The Store will be a generic dictionary (HashMap) which maps the identifier to a tuple of the target struct, and both sides of the channel we need to link up views to the object, and take action upon events. It will be implemented through a trait:

pub trait Relation<K: Copy + Eq + Hash, V: Clone + Default + Send> {
    fn get(id: K) -> V;
    fn insert(id: K, value: V);
    fn mutate(id: K, m: &impl Mutator<V>);
    fn subscribe(id: K) -> StateReceiver<V>;
    fn notify(id: K);
}

Implementing Relation<K, V> for V means that associated functions of V can be used to interact with uniquely identified instances of V - from anywhere in the DOM.

Mutations are a helper type that represent updates/changes to the struct as their own distinct struct. Even though it seems like unnecessary complexity, let this ride for a moment because it will become very useful when we want to pass those mutations to a remote back-end API, over the network.

Mutations just need to know how to apply themselves to their target type, producing a modified instance:

trait Mutator<V> {
    fn mutate(&self, v: &V) -> V;
}

An illustrative example of implementing Relation:

struct PostID(u32);

struct Post {
    id: PostID,
    body: String,
    author: String,
}

impl Relation<PostID, Post> for Post {
    // we need this impl to be easy...
}

// usage example:
let post = Post { 
    id: PostID(34), 
    body, 
    author, 
};
Post::insert(post.id, post);
let post = Post::get(PostID(34));

The branch in valerie (linked above), provides a macro implementation for such a dictionary (implemented using a HashMap). Let's have a look what that means for the TicTacToe code.

Abstracting the Model§ 

Firstly, we can abstract the Model. We can use some very simple and explicit structs.

A square is marked either empty, X or O, and starts out empty.

enum SquareMark {
    Empty,
    X,
    O,
}

impl Default for SquareMark {
    fn default() -> Self {
        Self::Empty
    }
}

We also use the mark to track which player is next (okok), so it needs a mutation which sets the starting player and changes between players.

enum NextPlayerChange {
    Start,
    Next,
}

impl Mutator<SquareMark> for NextPlayerChange {
    fn mutate(&self, v: &SquareMark) -> SquareMark {
        use SquareMark::*;
        match self {
            Self::Start => X,
            Self::Next => match v {
                X => O,
                O => X,
                Empty => unreachable!(),
            },
        }
    }
}

A square is state, indexable within the board. We set up some defaults so they start Empty, but I've omitted this code.

struct SquareID(u8);

struct Square {
    id: SquareID,
    pub mark: SquareMark,
}

Squares change when they are clicked, but the change depends on the player, so that is a parameter of the mutator.

enum SquareChange {
    Mark(SquareMark),
}

impl Mutator<Arc<Square>> for SquareChange {
    fn mutate(&self, v: &Arc<Square>) -> Arc<Square> {
        let mut v: Square = Square::clone(v);
        match self {
            Self::Mark(mark) => v.mark = *mark,
        }
        Arc::new(v)
    }
}

You will notice that we are mutating Arc<Square>, not Square. We pass data around a lot through channels, so our types need to be Clone.

The board is interesting. Instead of an array of marks, we have an array of identifiers. The Squares themselves are managed by their own relation, so we reference into that. Board no longer knows what constitutes a Square.

The default board sets up references to nine new squares, which in turn, default to Empty. We also factor out our game logic for determining if the board contains any lines of three identical non-Empty marks.

struct Board {
    pub squares: [SquareID; 9],
}

impl Default for Board {
    fn default() -> Self {
        let mut squares = [SquareID(0); 9];
        for i in 0usize..9 {
            let square = Square::new(SquareID(i as u8));
            Square::new(square.id);
            squares[i] = square.id;
        }
        Self { squares }
    }
}

impl Board {
    pub fn calculate_winner(&self) -> bool {
        // ...
    }
}

Game status has a very similar implementation to Square, and the mutator is the simplest setter, so not shown here.

Hopefully at this point you agree we have a model for playing a game of tictactoe, which is very much decoupled from the view.

I have not shown the implementations of the trait Relation yet. The code branch has two macros which implement the associated functions using a HashMap. All that is required to set up the Stores is:

// Square::get(SquareID) -> Arc<Square>
relation!(Square, SquareID, Arc<Square>);

// GameBoard::get() -> Arc<Board>
singleton!(GameBoard, Arc<Board>);
singleton!(GameStatus, Status);
singleton!(NextPlayer, SquareMark);

singleton! is very similar to relation! and sets up reactive state which is not indexed by an identifier.

We can represent our game state! What does it do for our view?

A clean view without model implementation noise§ 

It's time to get paid. But first, there is one part of the trait Relation I hid from you. formatted will take the instance and pass the changed object from the receiver channel to a closure which generates the Node text, which is inserted as the update.

fn game() -> Node {
    NextPlayer::mutate(NextPlayerChange::Start);
    execute(GameBoard::turn_checker());
    div!(div!(
        div!(
            GameStatus::formatted(move |s| {
                let s = match s {
                    Status::Playing => "Next player: ",
                    Status::Won => "Winner: ",
                };
                format!("{}", s)
            }),
            NextPlayer::formatted(move |p| {
                format!("{}", p)
            })
        )
        .class("status"),
        GameBoard::node()
    )
    .class("game-board"))
    .class("game")
    .into()
}

NextPlayer needed to be mutated because it is a singleton SquareMark and they default to Empty, remember?

game() executes an event-triggered function turn-checker, which we'll return to in a moment.

The game status is used to generate the label text, and the next player is used to display the mark of the next player. Finally, we insert the board as a node, which is a function we need to write, and we do it by re-opening the struct. It just loops through the squares in rows of three, and inserts three of the nodes created by square().

Notice that we are only passing the identifier of the square.

impl GameBoard {
    pub fn node() -> Node {
        let board = Self::get();
        let board = &board.squares;
        let mut parent = div!();
        for row in board.chunks(3) {
            let mut row_div: Tag<html::elements::Div> = div!().class("board-row");
            for id in row {
                row_div = row_div.push(square(*id));
            }
            parent = parent.push(row_div);
        }
        parent.into()
    }

    pub async fn turn_checker() {
        let rx = Self::subscribe();
        let mut old = StateId::new();
        while let Some((new, _)) = rx.receive(old).await {
            if GameBoard::get().calculate_winner() {
                GameStatus::mutate(StatusChange::Won);
            } else {
                NextPlayer::mutate(NextPlayerChange::Next);
            }
            old = new;
        }
    }
}

The turn-checker grabs a receiver for the GameBoard and waits for events. When an event arrives, it uses the model to check if we have a winner, and either ends the game, or flips the player, using the relevant mutators. The mutators for those stores push those changes down the channels so that the DOM is updated everywhere that is listening on related receivers.

fn square(id: SquareID) -> Node {
    button!(Square::formatted(id, |s| {
        format!("{}", s.mark)
    }))
    .class("square")
    .on_event("click", (), move |_, _| {
        use SquareMark::Empty;
        use Status::Playing;

        let status = GameStatus::get();
        let current = Square::get(id).mark;
        if status == Playing && current == Empty {
            Square::mutate(id, &SquareChange::Mark(NextPlayer::get()));
            GameBoard::notify();
        }
    })
    .into()
}

The square() function generates the buttons and sets up the event. It has a bit of logic around whether to accept the click, but if so, then it mutates the square to the current player's mark, and notifies the GameBoard. This triggers the turn-checker, which was looping, listening on the GameBoard's receiver.

Satisfy yourself now that if we put:

    GameBoard::node(),
    GameBoard::node(),
    GameBoard::node()

into game(), we will see three boards that are playing the same game. All three would run turn-checkers (which would be weird) but they will render the same square state from the static GameBoard singleton and Square relation.

Do you believe this is neater?§ 

We have done a lot of work to really separate the model and the view, and the views from each other. This is not quite enough return for what we have invested yet, but we have laid critical groundwork for very powerful back-end integration.

Next article, I'll explore some simple extensions to trait Relation that will let us fetch objects from a remote back-end, render them as loading... until they arrive, make local changes speculatively, and confirm them (or error) when the fetch returns to us with the server's outcome.