Bevy_save

A framework for saving and loading application state in Bevy.

https://user-images.githubusercontent.com/29737477/234151375-4c561c53-a8f4-4bfe-a5e7-b69af883bf65.mp4

Features

Save file management

bevy_save automatically uses your app's workspace name to create a unique, permanent save location in the correct place for any platform it can run on.

By default, World::save() and World::load() uses the managed save file location to save and load your application state, handling all serialization and deserialization for you.

Save directory location

With the default FileIO backend, your save directory is managed for you.

WORKSPACE is the name of your project's workspace (parent folder) name.

On WASM, snapshots are saved to LocalStorage, with the key WORKSPACE.KEY.

Reflection-based Snapshots

bevy_save is not just about save files, it is about total control over application state.

With the "reflect" feature enabled, this crate introduces a snapshot type which may be used directly:

• Snapshot is a serializable snapshot of all saveable resources, entities, and components.

Or via the World extension method WorldPathwayExt and WorldCheckpointExt:

• World::capture() captures a snapshot of the current application state, including resources.
• World::apply() applies a snapshot to the World.

Versioning and Migrations

Applications can change over the history of development. Users expect that saves created in an older version will continue to work in newer versions.

bevy_save provides support for reflection-based migrations with the Migrate trait:

// `#[reflect(Migrate)]` registers `ReflectMigrate` with the `TypeRegistry`
// This allows `Snapshot`s to save the type version and apply migrations automatically
#[derive(Reflect, Component, Debug, PartialEq)]
#[type_path = "migrate"]
#[type_name = "Position"]
#[reflect(Component, Migrate)]
struct Position {
    xyz: (f32, f32, f32),
}

// The `Migrate` trait allows you to define a `Migrator`
// which will step the upgrade through each version
impl Migrate for Position {
    fn migrator() -> Migrator<Self> {
        #[derive(Reflect)]
        #[type_path = "migrate"]
        #[type_name = "Pos"]
        struct PosV0_1 {
            x: f32,
            y: f32,
        }

        Migrator::new::<PosV0_1>("0.1.0")
            .version("0.2.0", |v1| {
                // Changing type paths and type names is supported
                #[derive(Reflect)]
                #[type_path = "migrate"]
                #[type_name = "Position"]
                struct PosV0_2 {
                    x: f32,
                    y: f32,
                }

                Some(PosV0_2 { x: v1.x, y: v1.y })
            })
            .version("0.3.0", |v2| {
                #[derive(Reflect)]
                #[type_path = "migrate"]
                #[type_name = "Position"]
                struct PosV0_3 {
                    x: f32,
                    y: f32,
                    z: f32,
                }

                // Fields can be re-mapped from version to version, added, or removed
                Some(PosV0_3 {
                    x: v2.x,
                    y: v2.y,
                    z: 0.0,
                })
            })
            // The final version will represent the current layout
            .version("0.4.0", |v2| {
                Some(Self {
                    xyz: (v2.x, v2.y, v2.z),
                })
            })
    }
}

Rollbacks and checkpoints

With the "checkpoints" feature enabled, this crate provides methods for creating checkpoints which are ordered and can be rolled back / forwards through.

• World::checkpoint() captures a snapshot for later rollback / rollforward.
• World::rollback() rolls the application state backwards or forwards through any checkpoints you have created.

The Checkpoints resource also gives you fine-tuned control of the currently stored rollback checkpoints.

Type registration

No special traits or NewTypes necessary, bevy_save takes full advantage of Bevy's built-in reflection. As long as the type implements Reflect, it can be registered and used with bevy_save.

bevy_save provides extension traits for App allowing you to do so.

• App.init_pipeline::<P>() initializes a Pipeline for use with save / load.
• App.allow_checkpoint::<T>() allows a type to roll back.
• App.deny_checkpoint::<T>() denies a type from rolling back.

Pipeline

The Pipeline trait allows you to use multiple different configurations of Backend and Format in the same App.

Using Pipeline also lets you re-use Snapshot appliers and builders.

struct HeirarchyPipeline;

impl Pipeline for HeirarchyPipeline {
    type Backend = DefaultDebugBackend;
    type Format = DefaultDebugFormat;

    type Key<'a> = &'a str;

    fn key(&self) -> Self::Key<'_> {
        "examples/saves/heirarchy"
    }

    fn capture(&self, builder: BuilderRef) -> Snapshot {
        builder
            .extract_entities_matching(|e| e.contains::<Player>() || e.contains::<Head>())
            .build()
    }

    fn apply(&self, world: &mut World, snapshot: &Snapshot) -> Result<(), bevy_save::Error> {
        snapshot
            .applier(world)
            .despawn::<Or<(With<Player>, With<Head>)>>()
            .apply()
    }
}

Type Filtering and Partial Snapshots

While bevy_save aims to make it as easy as possible to save your entire world, some applications also need to be able to save only parts of the world.

Builder allows you to manually create snapshots like DynamicSceneBuilder:

fn build_snapshot(world: &World, target: Entity, children: Query<&Children>) -> Snapshot {
    Snapshot::builder(world)
        // Extract all resources
        .extract_all_resources()

        // Extract all descendants of `target`
        // This will include all components not denied by the builder's filter
        .extract_entities(children.iter_descendants(target))

        // Entities without any components will also be extracted
        // You can use `clear_empty` to remove them
        .clear_empty()

        // Build the `Snapshot`
        .build()
}

Entity hooks

You are also able to add hooks when applying snapshots, similar to bevy-scene-hook.

This can be used for many things, like spawning the snapshot as a child of an entity:

let snapshot = Snapshot::from_world(world);

snapshot
    .applier(world)

    // This will be run for every Entity in the snapshot
    // It runs after the Entity's Components are loaded
    .hook(move |entity, cmds| {
        // You can use the hook to add, get, or remove Components
        if !entity.contains::<Parent>() {
            cmds.set_parent(parent);
        }
    })

    .apply();

Hooks may also despawn entities:

let snapshot = Snapshot::from_world(world);

snapshot
    .applier(world)

    .hook(|entity, cmds| {
        if entity.contains::<A>() {
            cmds.despawn();
        }
    })

Entity mapping

As Entity ids are not intended to be used as unique identifiers, bevy_save supports mapping Entity ids.

First, you'll need to get a ApplierRef:

• Snapshot::applier()
• ApplierRef::new()

The ApplierRef will then allow you to configure the entity map (and other settings) before applying:

let snapshot = Snapshot::from_world(world);

snapshot
    .applier(world)

    // Your entity map
    .entity_map(HashMap::default())

    // Despawn all entities matching (With<A>, Without<B>)
    .despawn::<(With<A>, Without<B>)>()

    .apply();

MapEntities

bevy_save also supports MapEntities via reflection to allow you to update entity ids within components and resources.

See Bevy's Parent Component for a simple example.

Flows

When creating a complex application, snapshot builder and applier functions tend to get complex and unwieldy.

Flows are chains of systems used to modularize this process, allowing you to build snapshots and apply them in stages.

They are defined similar to Bevy systems, but they require an input and an output.

Additionally, the introduction of Flows allows reflection to become optional - bring your own serialization if you so wish!

fn main() {
    App::new()
        .add_flows(CaptureFlow, (capture_tiles, capture_players, capture_cameras))
        .add_flows(ApplyFlow, (apply_tiles, apply_monsters));
}

// User-defined captures make reflection unnecessary
#[derive(Serialize, Deserialize)]
struct MyCapture {
    // ... but then you'll need to specify everything you extract
    tiles: Vec<(Entity, Tile)>,
    players: Vec<(Entity, Transform, Visibility)>,
    cameras: Vec<(Entity, Camera)>,
}

// Flow systems have full access to the ECS (even write access)
fn capture_tiles(In(cap): In<MyCapture>, tiles: Query<(Entity, &Tile)>) -> MyCapture {
    cap.tiles.extend(query.iter().map(|(e, t)| (e, t.clone())));
    cap
}

// Flow systems can be added to flows from anywhere, not just in one location
struct PluginA;

impl Plugin for PluginA {
    fn build(&self, app: &mut App) {
        app.add_flows(CaptureFlow, another_capture);
    }
}

Pathway

Pathway is the more flexible version of Pipeline which allows you to specify your own capture type and use Flows.

// Pathways look very similar to pipelines, but there are a few key differences
pub struct RONPathway;

impl Pathway for RONPathway {
    // The capture type allows you to save anything you want to disk, even without using reflection
    type Capture = Snapshot;

    type Backend = DefaultDebugBackend;
    type Format = RONFormat;
    type Key<'a> = &'a str;

    fn key(&self) -> Self::Key<'_> {
        "examples/saves/flows"
    }

    // Instead of capturing and applying directly, now these methods just return labels to user-defined flows
    // This allows for better dependency injection and reduces code complexity
    fn capture(&self, _world: &World) -> impl FlowLabel {
        CaptureFlow
    }

    fn apply(&self, _world: &World) -> impl FlowLabel {
        ApplyFlow
    }
}

// Flow labels don't encode any behavior by themselves, only point to flows
#[derive(Hash, Debug, PartialEq, Eq, Clone, Copy, FlowLabel)]
pub struct CaptureFlow;

#[derive(Hash, Debug, PartialEq, Eq, Clone, Copy, FlowLabel)]
pub struct ApplyFlow;

Backend

The Backend is the interface between your application and persistent storage.

Some example backends may include FileIO, sqlite, LocalStorage, or network storage.

#[derive(Default, Resource)]
pub struct FileIO;

impl<K: std::fmt::Display + Send> Backend<K> for FileIO {
    async fn save<F: Format, T: Serialize>(&self, key: K, value: &T) -> Result<(), Error> {
        let path = get_save_file(format!("{key}{}", F::extension()));
        let dir = path.parent().expect("Invalid save directory");
        create_dir_all(dir).await?;
        let mut buf = Vec::new();
        F::serialize(&mut buf, value)?;
        let mut file = File::create(path).await?;
        Ok(file.write_all(&buf).await?)
    }

    async fn load<F: Format, S: for<'de> DeserializeSeed<'de, Value = T>, T>(
        &self,
        key: K,
        seed: S,
    ) -> Result<T, Error> {
        let path = get_save_file(format!("{key}{}", F::extension()));
        let mut file = File::open(path).await?;
        let mut buf = Vec::new();
        file.read_to_end(&mut buf).await?;
        F::deserialize(&*buf, seed)
    }
}

Format

Format is how your application serializes and deserializes your data.

Formats can either be human-readable like JSON or binary like MessagePack.

pub struct RONFormat;

impl Format for RONFormat {
    fn extension() -> &'static str {
        ".ron"
    }

    fn serialize<W: Write, T: Serialize>(writer: W, value: &T) -> Result<(), Error> {
        let mut ser = ron::Serializer::new(
            writer.write_adapter(),
            Some(ron::ser::PrettyConfig::default()),
        )
        .map_err(Error::saving)?;

        value.serialize(&mut ser).map_err(Error::saving)
    }

    fn deserialize<R: Read, S: for<'de> DeserializeSeed<'de, Value = T>, T>(
        reader: R,
        seed: S,
    ) -> Result<T, Error> {
        ron::options::Options::default()
            .from_reader_seed(reader, seed)
            .map_err(Error::loading)
    }
}

Prefabs

The Prefab trait allows you to easily spawn entities from a blueprint.

#[derive(Component, Default, Reflect)]
#[reflect(Component)]
struct Ball;

#[derive(Reflect)]
struct BallPrefab {
    position: Vec3,
}

impl Prefab for BallPrefab {
    type Marker = Ball;

    fn spawn(self, target: Entity, world: &mut World) {
        // Some entities will need initialization from world state, such as mesh assets.
        // We can do that here.
        let mesh = world.resource_mut::<Assets<Mesh>>().add(Circle::default());
        let material = world
            .resource_mut::<Assets<ColorMaterial>>()
            .add(BALL_COLOR);

        world.entity_mut(target).insert((
            Mesh2d(mesh),
            MeshMaterial2d(material),
            Transform::from_translation(self.position)
                .with_scale(Vec2::splat(BALL_DIAMETER).extend(1.)),
            Ball,
            Velocity(INITIAL_BALL_DIRECTION.normalize() * BALL_SPEED),
        ));
    }

    fn extract(builder: BuilderRef) -> BuilderRef {
        // We don't actually need to save all of those runtime components.
        // Only save the translation of the Ball.
        builder.extract_prefab(|entity| {
            Some(BallPrefab {
                position: entity.get::<Transform>()?.translation,
            })
        })
    }
}

You are also able to extract resources by type:

Snapshot::builder(world)
    // Extract the resource by the type name
    // In this case, we extract the resource from the `manual` example
    .extract_resource::<FancyMap>()

    // Build the `Snapshot`
    // It will only contain the one resource we extracted
    .build()

Additionally, explicit type filtering like ApplierRef is available when building snapshots:

Snapshot::builder(world)
    // Exclude `Transform` from this `Snapshot`
    .deny::<Transform>()

    // Extract all matching entities and resources
    .extract_all()

    // Clear all extracted entities without any components
    .clear_empty()

    // Build the `Snapshot`
    .build()

Stability

bevy_save attempts to provide stability guarantees for Snapshot serialization and deserialization between crate versions on a best-effort basis, enforced with unit tests.

If a breaking change is introduced, the number after the + on the crate version will be incremented and it will be supported via the version method on SnapshotDeserializer.

bevy_save relies on serialization to create save files and as such is exposed to internal implementation details for types. As a result, Bevy or other crate updates may break your save file format. It should be possible to mitigate this by defining ReflectMigrate for any offending types.

Changing what entities have what components or how you use your entities or resources in your logic can also result in broken saves.

Entity

For all intents and purposes, Entity should be treated as an opaque identifier. The internal bit representation is liable to change from release to release as are the behaviors or performance characteristics of any of its trait implementations (i.e. Ord, Hash, etc.). This means that changes in Entity’s representation, though made readable through various functions on the type, are not considered breaking changes under SemVer.

In particular, directly serializing with Serialize and Deserialize make zero guarantee of long term wire format compatibility. Changes in behavior will cause serialized Entity values persisted to long term storage (i.e. disk, databases, etc.) will fail to deserialize upon being updated.

— Bevy's Entity documentation

bevy_save serializes and deserializes entities directly. If you need to maintain compatibility across Bevy versions, consider adding a unique identifier Component to your tracked entities.

Compatibility

bevy_save follows Semantic Versioning 2.0.0, with additional metadata: MAJOR.MINOR.PATCH+SNAPSHOT

• Major: Breaking API changes and/or Snapshot format changes
• Minor: Backwards-compatible API changes
• Patch: Backwards-compatible bug-fixes
• Snapshot: Snapshot version, incremented when the wire format of Snapshot changes in a way that will break existing applications

Bevy

Snapshot Version

0. bevy_save introduced serialization support
1. bevy_save introduced a new Snapshot format
2. bevy changed Entity's on-disk representation
3. bevy_save began using FromReflect when taking snapshots
4. bevy_save introduced a new Snapshot format, see below

Migrating

Platforms

Feature Flags

License

bevy_save is dual-licensed under MIT and Apache-2.0.