A framework for saving and loading application state in Bevy.
preview-0.6.0.mp4
When creating a complex application, snapshot builder and applier functions tend to get complex and unwieldy.
Flow
s 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 Flow
s 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
is the more flexible version of Pipeline
which allows you to specify your own capture type and use Flow
s.
// 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;
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.
With the default FileIO
backend, your save directory is managed for you.
WORKSPACE
is the name of your project's workspace (parent folder) name.
Windows | Linux/*BSD | MacOS |
---|---|---|
C:\Users\%USERNAME%\AppData\Local\WORKSPACE\saves |
~/.local/share/WORKSPACE/saves |
~/Library/Application Support/WORKSPACE/saves |
On WASM, snapshots are saved to LocalStorage
, with the key WORKSPACE.KEY
.
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 theWorld
.
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),
})
})
}
}
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.
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 aPipeline
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.
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()
}
}
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()
}
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();
}
})
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
:
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();
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.
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
is how your application serializes and deserializes your data.
Format
s 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)
}
}
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()
bevy_save
attempts to provide stability guarantees for Snapshot
serialization and deserialization between crate versions, enforced with unit tests.
If a breaking change is introduced, 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.
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 inEntity
’s representation, though made readable through various functions on the type, are not considered breaking changes under SemVer.In particular, directly serializing with
Serialize
andDeserialize
make zero guarantee of long term wire format compatibility. Changes in behavior will cause serializedEntity
values persisted to long term storage (i.e. disk, databases, etc.) will fail to deserialize upon being updated.
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.
Bevy Version | Crate Version |
---|---|
0.16 |
0.18 , 0.19 , 1.0 3 |
0.15 |
0.16 2, 0.17 |
0.14 1 |
0.15 |
0.13 |
0.14 |
0.12 |
0.10 , 0.11 , 0.12 , 0.13 |
0.11 |
0.9 |
0.10 |
0.4 , 0.5 , 0.6 , 0.7 , 0.8 |
0.9 |
0.1 , 0.2 , 0.3 |
-
bevy
changedEntity
's on-disk representation -
bevy_save
began usingFromReflect
when taking snapshots -
bevy_save
removed thecheckpoints
field fromSnapshot
, instead savingCheckpoints
as a resource viaReflect
. Use theversion
method onSnapshotDeserializer
to load a snapshot created in the previous version. Types implementingMigrate
will deserialize from snapshots created in the previous version automatically.
This version introduced Pathway
, which is effectively a superset of Pipeline
.
- In
World::capture
,World:apply
,World::save
,World::load
methods and similar, add a&
before your existing pipeline - Previously provided
Commands
extension traits and associated commands have been removed (sincePathway
operates on references), you'll need to write your own or use events instead - If you're using
default-features = false
, you'll need to add thereflect
andcheckpoints
features in order to get parity with the last version SnapshotBuilder
andSnapshotApplier
have been renamed toBuilderRef
andApplierRef
, respectively.
This version introduced versioning and migrations.
SnapshotVersion::V0_16
can be used with theversion
method onSnapshotDeserializer
to load a snapshot created in a previous version if the snapshot had checkpoints.- Snapshots created in a previous version without checkpoints should load as expected.
- The fields for all serializers and deserializers have been made private. Use the
new
methods to construct them. - Non self-describing formats such as
postcard
should now work as expected.
Platform | Support |
---|---|
Windows | Yes |
MacOS | Yes |
Linux | Yes |
WASM | Yes |
Android | No |
iOS | No |
Feature flag | Description | Default? |
---|---|---|
reflect |
Enables reflection-based snapshots | Yes |
checkpoints |
Enables reflection-based checkpoints | Yes |
bevy_asset |
Enables bevy_asset type registration |
Yes |
bevy_render |
Enables bevy_render type registration |
Yes |
bevy_sprite |
Enables bevy_sprite type registration |
Yes |
brotli |
Enables Brotli compression middleware |
No |
bevy_save
is dual-licensed under MIT and Apache-2.0.