It’s been about 6 months since I watched Catherine West’s excellent Using Rust for Game Development sent me down the Entity-Component-System (ECS) rabbit hole, and I thought I’d share some of my findings.

I’ve been meaning to write about this for quite a while now but it took a while to put my thoughts into a cohesive article without throwing massive walls of code at you.

This article is mainly focused around the high-level decisions you make when designing a project, so there won’t be as much code as normal. That said, all code written in this article is available on GitHub. Feel free to browse through and steal code or inspiration.

If you found this useful or spotted a bug, let me know on the blog’s issue tracker!

What Is An Entity-Component-System? Link to heading

I hope you’ll forgive a little copy-paste, but the Wikipedia definition gives a fairly decent summary of the topic:

ECS follows the composition over inheritance principle that allows greater flexibility in defining entities where every object in a game’s scene is an entity (e.g. enemies, bullets, vehicles, etc.). Every entity consists of one or more components which add behavior or functionality. Therefore, the behavior of an entity can be changed at runtime by adding or removing components.

  • Entity: The entity is a general purpose object. Usually, it only consists of a unique id. They “tag every coarse gameobject as a separate item”. Implementations typically use a plain integer for this.
  • Component: The raw data for one aspect of the object, and how it interacts with the world. “Labels the Entity as possessing this particular aspect”. Implementations typically use structs, classes, or associative arrays.
  • System: “Each System runs continuously (as though each System had its own private thread) and performs global actions on every Entity that possesses a Component of the same aspect as that System.”

There are several high-quality ECS implementations, but specs crate is widely accepted as one of the best ECS libraries in Rust.

For me an ECS is an architectural pattern for data-heavy applications which enforces a clear distinction between behaviour and data, and embodies the Composition over Inheritance way of doing things.

Inheritance isn’t Always the Best Tool for the Job Link to heading

The inspiration for trying the ECS architecture outside of games comes from one of my work projects.

Without going into too much detail, the CAD/CAM program we’ve written at work is built on top of a 3rd party CAD engine. This CAD engine is a native library which exposes a heavily object-oriented interface, and there are a quite a few places where the inadequacies of structuring everything around inheritance show through.

CAD libraries are composed of many cross-cutting concerns, you have:

  • Graphical entities (e.g. Point, Line, Spline) which are rendered to the screen
  • Non-graphical entities which impart semantics to the drawing (e.g. graphical entities can be grouped into Layers which be individually managed)
  • Both graphical and non-graphical entities can be frozen (made immutable) or hidden (made invisible)
  • Entities (both graphical and non-graphical) can be given a name so users associate them with a concept (e.g. you may put all dimension lines on the "dimensions" layer), allowing easy lookup and letting the UI differentiate between different entities of the same type
  • Different graphical entities need different information for how to be rendered (e.g. a Line might just have a stroke_colour, while a Circle may also have a fill_colour)

Trying to model all of this using a typical object-oriented architecture is really tricky.

Let’s say you create a GraphicalEntity base class for all types which can be rendered to the screen. Types like Line, Circle, and Spline are all drawn using lines so it makes sense to have some stroke property (imagine it contains the line width and colour). Instead of adding the stroke property to all three classes individually, you decide to pull the property up into the parent class to avoid duplication and let us set some_graphical_entity.stroke without caring whether it is a Line or a Circle or a Spline.

But this introduces a bit of a problem. We want to display images on a drawing so there’s an Image class which inherits from GraphicalEntity. However an Image is drawn completely differently to a Line or Circle, so the stroke property we introduced earlier is just bloat.

classDiagram GraphicalEntity <|-- Line GraphicalEntity <|-- Spline GraphicalEntity <|-- Circle GraphicalEntity <|-- Image GraphicalEntity: Stroke stroke GraphicalEntity: void render() Circle: Colour fill_colour Image: byte[] pixel_buffer

Okay, so maybe our original inheritance hierarchy adds some unnecessary bloat but it’s not the end of the world… right?

We also want to draw points, zero-dimension dots on the drawing. We can just add a Point class which inherits from GraphicalEntity. So far, so good.

It’d be nice to have a function for decomposing complex graphical entities into simpler ones (e.g. to approximate a Spline using arcs and lines). We can’t give the GraphicalEntity class some decompose() method because that just wouldn’t make sense for Image, so let’s introduce an intermediate DecomposeableEntity.

classDiagram GraphicalEntity <|-- DecomposeableEntity GraphicalEntity <|-- Image GraphicalEntity: Stroke stroke GraphicalEntity: void render() DecomposeableEntity <|-- Line DecomposeableEntity <|-- Spline DecomposeableEntity <|-- Circle class DecomposeableEntity { GraphicalEntity[] decompose() } Circle: Colour fill_colour Image: byte[] pixel_buffer

While we’re at it we also want to have hatching, a common drafting technique used to show which areas of a drawing are part of the same thing. Hatches are really just a set of diagonal lines, so it makes sense that the class should inherit from DecomposeableEntity. It’s not uncommon for hatching to colour the background a different colour, so let’s give it a fill_colour property.

But hang on… doesn’t Circle also have a fill_colour property? What if we DRY things up by creating a new class called DecomposeableEntityWithFill?

classDiagram GraphicalEntity <|-- DecomposeableEntity GraphicalEntity <|-- Image GraphicalEntity: Stroke stroke GraphicalEntity: void render() DecomposeableEntity <|-- Line DecomposeableEntity <|-- Spline DecomposeableEntity <|-- DecomposeableEntityWithFill DecomposeableEntityWithFill <|-- Circle DecomposeableEntityWithFill <|-- Hatch class DecomposeableEntity { GraphicalEntity[] decompose() } Image: byte[] pixel_buffer DecomposeableEntityWithFill: Colour fill_colour

The diagonal lines in a Hatch don’t actually exist on the drawing though. Instead they’re rendered a fixed distance apart regardless of the zoom level, so we’ll also need to add a zoom_level parameter to the decompose() function. It’s a little annoying because things like Line and Circle don’t actually care about how far we’re zoomed in, but we’re already bloating the GraphicalEntity class with unused properties like stroke so what harm will a little extra bloat do?

You can see where this is going. For every new property we could try to reuse code by introducing intermediate classes, but it won’t be long before we code ourselves into a corner. Unfortunately, the real world doesn’t fit into a tidy inheritance hierarchy.

It’s not long before your class hierarchy is ten levels deep, bloated with loads of unnecessary data and methods, and there are so many levels of “abstraction” it’s hard to figure out what’s actually going on.

Another problem is you’ll frequently fall into the Refused Bequest anti-pattern. This is where a parent class exposes a method that doesn’t actually make sense for some child classes so the child class overrides it to always throw a throw new InvalidOperationException(). Everything still compiles, but now every time you invoke the method on the parent class there’ll be a niggling feeling in the back of your head that things may blow up at runtime.

That’s not a fun feeling. Especially when you’re letting your project manager demo the application and he starts experimenting with combinations of operations you never anticipated or tested for… Don’t ask me how I know this 😑

As an aside, have you ever heard of the Circle-Ellipse Problem?

If we have an application that uses circles and ellipses (e.g. a graphics program), should we have two classes Circle and Ellipse? Which should inherit from which, if at all? A circle is a special kind of ellipse, viz. one where the two foci coincide. But if an Ellipse is mutable, a Circle is mutable too, and can be made a non-circle.

Or should we only have an Ellipse? But if we then create an Ellipse that happens to represent a circle, we cannot ask it for its radius, because Ellipse has no radius() method.

Most object-oriented languages are designed so that an object’s underlying type will be the same for its entire lifetime. This makes things interesting when users want to scale a Circle without maintaining aspect ratio. It means you can’t just give the GraphicalEntity a scale() method which mutates the object in-place, you need to change the entire API so a Circle can return an Ellipse when the x and y scale factors aren’t the same.

If you’ve been programming for a while you will have probably come across the mantra, “Composition over Inheritance”. It’s exactly these sorts of design problems composition is attempting to solve, and ECS is just one way to formalise composition… By breaking the world up into Components (data that can be attached to things) and Systems (behaviour).

Creating an ECS-based CAD Library Link to heading

I’m a big fan of the specs crate, so that’s what I used when trying to implement an ECS-based CAD library.

I’m also really boring when it comes to naming things, so the project is simply called A Rust CAD System, or arcs for short. This is also a nice pun on the fact that one of the basic drawing primitives of any CAD library is the Arc 😁

All graphical entities have a DrawingObject component which contains the data which is needed while rendering.

// arcs/src/components/drawing_object.rs

/// Something which can be drawn on the screen.
#[derive(Debug, Clone, PartialEq)]
pub struct DrawingObject {
    pub geometry: Geometry,
    /// The [`Layer`] this [`DrawingObject`] is attached to.
    pub layer: Entity,
}

impl Component for DrawingObject {
    type Storage = FlaggedStorage<Self, DenseVecStorage<Self>>;
}

/// The geometry of a [`DrawingObject`].
#[derive(Debug, Clone, PartialEq)]
#[non_exhaustive]
pub enum Geometry {
    Line(Line),
    Arc(Arc),
    Point(Point),
    ...
}

You may have noticed that we’re explicitly implementing Component for DrawingObject instead of using the custom derive. This is because we want to store this component using FlaggedStorage, a wrapper type which lets you subscribe to change notifications.

You’ll see why change notifications are useful later on.

Rendering Link to heading

I’m using the piet crate as an abstraction over a drawing canvas. This is awesome because not only has all the hard work been implemented, including tricky things like fonts and gradients, but there are also backends for all the major platforms. Including the browser. This means we can create a an online demo later on by compiling to WebAssembly, which is a massive boon when trying to show other people your work… It’s also just a well-written library and does exactly what I need.

The piet-web backend introduces a minor complication (in the form of mental overhead) because its RenderContext borrows JavaScript objects. That means every time we need to render we’ll have to create a temporary System which holds a reference to a particular piet backend, instead of implementing System on the Renderer directly.

// arcs/render/renderer.rs

/// Long-lived state used when rendering.
#[derive(Debug, Clone)]
#[non_exhaustive]
pub struct Renderer {
    pub viewport: Viewport,
    pub background: Color,
}

impl Renderer {
    pub fn new(viewport: Viewport, background: Color) -> Self {
        Renderer {
            viewport,
            background,
        }
    }

    /// Get a [`System`] which will render using a particular [`RenderContext`].
    pub fn system<'a, R>(
        &'a mut self,
        backend: R,
        window_size: Size,
    ) -> impl System<'a> + 'a
    where
        R: RenderContext + 'a,
    {
        RenderSystem { backend, window_size, renderer: self }
    }
}

#[derive(Debug, Clone, PartialEq)]
pub struct Viewport {
    /// The location (in drawing units) this viewport is centred on.
    pub centre: Vector,
    /// The number of pixels each drawing unit should take up on the screen.
    pub pixels_per_drawing_unit: f64,
}

/// The [`System`] which actually renders things.
///
/// This needs to be a temporary object "closing over" the [`Renderer`] and some
/// [`RenderContext`] due to lifetimes.
///
/// In particular, the `RenderContext` for the `piet_web` crate takes the HTML5
/// canvas by `&mut` reference instead of owning it, and we don't want to tie our
/// [`Renderer`] to a particular stack frame because it's so long lived (we'd end
/// up fighting the borrow checker and have self-referential types).
#[derive(Debug)]
struct RenderSystem<'renderer, B> {
    backend: B,
    window_size: Size,
    renderer: &'renderer mut Renderer,
}

Going through the entire rendering system is out of scope for this article, but I’ll walk you through how we use specs Components to nicely manage things like the different styling information attached to the various graphical entities.

The RenderSystem’s System impl is surprisingly simple. We break the task up into calculating the draw order (the user can specify that certain objects should be drawn on top of others) and then iterating through each entity to be drawn and calling self.render() on them.

// arcs/render/renderer.rs

impl<'world, 'renderer, B: RenderContext> System<'world>
    for RenderSystem<'renderer, B>
{
    type SystemData = (DrawOrder<'world>, Styling<'world>);

    fn run(&mut self, data: Self::SystemData) {
        // make sure we're working with a blank screen
        self.backend.clear(self.renderer.background.clone());

        let (draw_order, styling) = data;

        let viewport_dimensions = self.viewport_dimensions();

        for (ent, obj) in draw_order.calculate(viewport_dimensions) {
            self.render(ent, obj, &styling);
        }
    }
}

We’ve created helper struct called DrawOrder which holds a reference to each set of Components we’ll need while calculating the draw order.

// arcs/src/render/renderer.rs

/// The state needed when calculating which order to draw things in so z-levels
/// are implemented correctly.
#[derive(SystemData)]
struct DrawOrder<'world> {
    entities: Entities<'world>,
    drawing_objects: ReadStorage<'world, DrawingObject>,
    layers: ReadStorage<'world, Layer>,
    bounding_boxes: ReadStorage<'world, BoundingBox>,
}

impl<'world> DrawOrder<'world> {
    fn calculate(
        &self,
        viewport_dimensions: BoundingBox,
    ) -> impl Iterator<Item = (Entity, &'_ DrawingObject)> + '_ {
        type EntitiesByZLevel<'a> =
            BTreeMap<Reverse<usize>, Vec<(Entity, &'a DrawingObject)>>;

        // Iterate through all drawing objects, grouping them by the parent
        // layer's z-level in reverse order (we want to yield higher z-levels
        // first)
        let mut drawing_objects = EntitiesByZLevel::new();

        // PERF: This function has a massive impact on render times
        // Some ideas:
        //   - Use a pre-calculated quad-tree so we just need to check items
        //     within the viewport bounds
        //   - use a entities-to-layers cache so we can skip checking whether to
        //     draw an object on a hidden layer

        for (ent, obj, bounds) in (
            &self.entities,
            &self.drawing_objects,
            MaybeJoin(&self.bounding_boxes),
        )
            .join()
        {
            let Layer { z_level, visible } = self
                .layers
                .get(obj.layer)
                .expect("The object's layer was deleted");

            // try to use the cached bounds, otherwise re-calculate them
            let bounds = bounds
                .copied()
                .unwrap_or_else(|| obj.geometry.bounding_box());

            if *visible && viewport_dimensions.intersects_with(bounds) {
                drawing_objects
                    .entry(Reverse(*z_level))
                    .or_default()
                    .push((ent, obj));
            }
        }

        drawing_objects.into_iter().flat_map(|(_, items)| items)
    }
}

It’s not uncommon for a drawing to contain hundreds of thousands of graphical entities, so it’s really important to reduce the amount of work that gets done. You can see from the PERF comment that we’re willing to trade off extra memory usage if it means we can reduce the rendering system’s execution time.

Let me know if you can see possible bugs or other improvements by making an issue against the project’s issue tracker. I’m especially keen to hear if you’ve had to tackle these sorts of problems before!

When rendering a Point, there are a couple pieces of information we’ll need. These are stored using the PointStyle component.

// arcs/components/styles.rs

#[derive(Debug, Clone, Component)]
#[storage(DenseVecStorage)]
pub struct PointStyle {
    pub colour: Color,
    pub radius: Dimension,
}

impl Default for PointStyle {
    fn default() -> PointStyle {
        PointStyle {
            colour: Color::BLACK,
            radius: Dimension::default(),
        }
    }
}

/// A dimension on the canvas.
#[derive(Debug, Copy, Clone, PartialEq)]
pub enum Dimension {
    /// The dimension should always be the same size in pixels, regardless of
    /// the zoom level.
    Pixels(f64),
    /// A "real" dimension defined in *Drawing Space*, which should be scaled
    /// appropriately when we zoom.
    DrawingUnits(f64),
}

impl Dimension {
    pub fn in_pixels(self, pixels_per_drawing_unit: f64) -> f64 {
        match self {
            Dimension::Pixels(px) => px,
            Dimension::DrawingUnits(units) => units * pixels_per_drawing_unit,
        }
    }
}

impl Default for Dimension {
    fn default() -> Dimension { Dimension::Pixels(1.0) }
}

Rendering a Point then becomes a process of:

  1. Find the PointStyle to use for this entity
  2. Define a kurbo::Shape for the point’s outline, in this case a Circle
  3. Tell the backend to fill the Circle with the desired colour
// arcs/src/render/renderer.rs

impl<'world, 'renderer, B: RenderContext> RenderSystem<'renderer, B> {
    fn render(
        &mut self,
        ent: Entity,
        drawing_object: &DrawingObject,
        styles: &Styling,
    ) {
        match drawing_object.geometry {
            Geometry::Point(ref point) => {
                self.render_point(ent, point, drawing_object.layer, styles)
            },

            ...
        }
    }

    /// Draw a [`Point`] as a circle on the canvas.
    fn render_point(
        &mut self,
        entity: Entity,
        point: &Point,
        layer: Entity,
        styles: &Styling,
    ) {
        let fallback = PointStyle::default();

        let style = styles
            .point_styles
            // the style for this point may have been overridden explicitly
            .get(entity)
            // otherwise fall back to the layer's PointStyle
            .or_else(|| styles.point_styles.get(layer))
            // fall back to the global default if the layer didn't specify one
            .unwrap_or(&fallback);

        let point = Circle {
            center: self.to_viewport_coordinates(point.location),
            radius: style
                .radius
                .in_pixels(self.renderer.viewport.pixels_per_drawing_unit),
        };

        self.backend.fill(point, &style.colour);
    }

    /// Translates a [`Vector`] from drawing space to a [`kurbo::Point`] on the
    /// canvas.
    fn to_viewport_coordinates(&self, point: Vector) -> kurbo::Point {
        super::to_canvas_coordinates(
            point,
            &self.renderer.viewport,
            self.window_size,
        )
    }
}

Bounding Boxes Link to heading

To make sure we only try to draw things within the rendering system’s viewport each graphical object is given an axis-aligned BoundingBoxes component. To avoid needing to remember to update this BoundingBox component every time an object is updated we can make use of the DrawingObject’s FlaggedStorage and create a SyncBounds system which will subscribe to changes and ensure object bounds are kept in sync.

The SyncBounds implementation is copied almost directly from the docs for FlaggedStorage.

// arcs/systems/bounds.rs

/// Lets us keep track of a [`DrawingObject`]'s rough location in *Drawing
/// Space*.
#[derive(Debug)]
pub struct SyncBounds {
    changes: ReaderId<ComponentEvent>,
    to_update: BitSet,
    removed: BitSet,
}

impl SyncBounds {
    pub const NAME: &'static str = module_path!();

    pub fn new(world: &World) -> SyncBounds {
        SyncBounds {
            changes: world.write_storage::<DrawingObject>().register_reader(),
            to_update: BitSet::new(),
            removed: BitSet::new(),
        }
    }
}

impl<'world> System<'world> for SyncBounds {
    type SystemData = (
        WriteStorage<'world, BoundingBox>,
        ReadStorage<'world, DrawingObject>,
        Entities<'world>,
    );

    fn run(&mut self, data: Self::SystemData) {
        // clear any left-over flags
        self.to_update.clear();
        self.removed.clear();

        let (mut bounds, drawing_objects, entities) = data;

        // find out which items have changed since we were last polled
        for event in drawing_objects.channel().read(&mut self.changes) {
            match *event {
                ComponentEvent::Inserted(id) | ComponentEvent::Modified(id) => {
                    self.to_update.add(id);
                },
                ComponentEvent::Removed(id) => {
                    self.removed.add(id);
                },
            }
        }

        for (ent, drawing_object, _) in
            (&entities, &drawing_objects, &self.to_update).join()
        {
            bounds
                .insert(ent, drawing_object.geometry.bounding_box())
                .unwrap();
        }

        for (ent, _) in (&entities, &self.removed).join() {
            bounds.remove(ent);
        }
    }
}

We may also want to override the System::setup() method to go through all DrawingObject entities and make sure they’ve got a BoundingBox component.

In general, if we ever need to cache something we’ll create one of these bookkeeping Systems. We can take advantage of the DispatcherBuilder to register any necessary bookkeeping tasks with a Dispatcher using a function defined in the systems module.

// arcs/systems/mod.rs

/// Register any necessary background tasks with a [`DispatcherBuilder`].
pub fn register_background_tasks<'a, 'b>(
    builder: DispatcherBuilder<'a, 'b>,
    world: &World,
) -> DispatcherBuilder<'a, 'b> {
    builder.with(SyncBounds::new(world), SyncBounds::NAME, &[])
}

Conclusion Link to heading

As far as I can tell, using an ECS for managing the data in a CAD library seems to work pretty well. I’m thinking of building an online editor for Ladder Logic programs (specs can be compiled to WebAssembly without a problem), so I’ll hopefully make another article later on telling you how things go.

I’ve also experimented with using specs as the backend for a compiler in the past. When writing a compiler you often end up implementing a poor man’s ECS anyway (i.e. IR nodes are entities, each “pass” is a System, and the various side-tables and metadata can be attached to IR nodes as Components) so from a theoretical perspective using a proper ECS in a compiler makes a lot of sense.

Once I’ve got a basic editor for Ladder Logic programs I’m planning to revisit this way idea when compiling programs to an executable form (e.g. WebAssembly).

See Also: