bevy_timewarp

Buffer and rollback to states up to a few frames ago, for rollback networking.

Doesn't do any networking, just concerned with buffering old component states in order to revert values to previous frames and quickly fast-forward back to the original frame. Assumes game logic uses bevy's FixedUpdate schedule.

Current status: under development alongside a multiplayer server-authoritative game. I am "RJ" on bevy's discord, in the networking channel, if you want to discuss.

"Sports Games"

This crate is built for so called "sports games" where dynamic entities interact in the same timeframe. That means clients simulate all entities into the future, in the same timeframe as the local player. This necessitates predicting player inputs.

This is quite different to traditional Quake style FPS netcode, which is for a mostly static world, with players running around shooting eachother. In Quake style, you are always seeing a delayed version of other players, interpolated between two snapshots. The server does backwards reconcilliation / lag compensation to verify hits.

Example rollback scenario:

In your client/server game:

• client is simulating frame 10
• server snapshot for frame 6 arrives, including values for an entity's component T
• client updates entity's ServerSnapshot value at frame 6 (ie, the past)
• Timewarp triggers a rollback to frame 6 if snapshot != our stored value for frame 6.
• • winds back frame counter to 6
  • copies the server snapshot value to the component
  • resimulates frames 7,8,9,10 as fast as possible
  • during this process, your systems would apply player inputs you've stored for each frame
• Rollback ends and frame 11 continues normally.

Quickstart

Add the plugin, then register the components you wish to be rollback capable:

// Store 10 frames of rollback, run timewarp systems after our GameLogic set.
app.add_plugins(TimewarpPlugin::new(10, MySets::GameLogic));
// register components that should be buffered and rolled back as needed:
app.register_rollback::<MyComponent>();
app.register_rollback::<Position>();
// etc..

Any entity that has a T Component will automatically be given a [ComponentHistory<T>] and [ServerSnapshot<T>] component.

ComponentHistory is a circular buffer of the last N frames of component values. This is logged every frame automatically, so is mostly your client predicted values. You typically won't need to interact with this.

ServerSnapshot is a buffer of the last few authoritative component values, typically what you received from the game server. Your network system will need to add new values to this.

When you receive authoritative updates, add them to the ServerSnapshot like so:

fn process_position_updates_from_server(
    mut q: Query<(Entity, &mut ServerSnapshot<Position>)>,
    update: Res<UpdateFromServer>, // however you get your updates, not our business.
){
    // typically update.frame would be in the past compared to current client frame
    for (entity, mut ss_position) in q.iter_mut() {
        let component_val = update.get_position_at_frame_for_entity(update.frame, entity); // whatever
        ss_position.insert(update.frame, component_val);
    }
}

Alternatively, and especially if you are inserting a component your entity has never had before, meaning there will be no ServerSnapshot<T> component, insert components in the past like this:

let add_at_frame = 123;
let historical_component = InsertComponentAtFrame::new(add_at_frame, MyComponent);
commands.entity(e1).insert(historical_component);

Systems configuration

Divide up your game systems so that during a rollback you still apply stored player input, but ignore stuff like sending network messages etc.

During a rollback, the [Rollback] resource will exist. Use this in a run_if condition.

// Normal game loop when not doing a rollback/fast-forward
app.add_systems(FixedUpdate,
    (
        frame_inc,
        process_server_messages,
        process_position_updates_from_server,
        spawn_stuff,
        read_player_inputs,
        apply_all_player_inputs_to_simulation_for_this_frame,
        do_physics,
        render,
        etc,
    )
    .chain()
    .in_set(MySets::GameLogic)
    .run_if(not(resource_exists::<Rollback>())) // NOT in rollback
);
// Abridged game loop for replaying frames during rollback/fast-forward
app.add_systems(FixedUpdate,
    (
        frame_inc,
        apply_all_player_inputs_to_simulation_for_this_frame,
        do_physics,
    )
    .chain()
    .in_set(MySets::GameLogic)
    .run_if(resource_exists::<Rollback>()) // Rollback is happening.
);

Visual smoothing of errors

Timewarp snaps the simulation state – ie. the value of a component at a specific frame simulated locally vs the value after rollback and resimulate might differ. If you register your components like this:

    app.register_rollback_with_correction_logging::<Position>();

then timewarp will capture the before and after versions of components when doing a rollback, and put it into a [TimewarpCorrection<Position>] component for your game to examine. Typically this would be useful for some visual smoothing - you might gradually blend over the error distance with your sprite, even though the underlying physical simulation snapped correct.

Testing various edge cases

TODO: I don't know how to link rustdocs to integration tests..

See the basic_rollback test for the most straightforward scenario: a long lived entity with components that haven't been added/removed, and an authoritative server update arrives. Apply value to past frame, rollback and continue.

The despawn_markers test illustrates how to despawn – rather than doing commands.entity(id).despawn(), which would remove all trace and thus mean the entity couldn't be revived if we needed to rollback to a frame when it was alive, you insert a [DespawnMarker(frame_number)][DespawnMarker] component, which cleans up the entity immediately by removing all its registered components, then does the actual despawn after rollback_window frames have elapsed.

The despawn_revival_during_rollback test does something similar, but triggers a rollback which will restore components to an entity tagged with a DespawnMarker in order to resimulate after a server update arrives, and then remove the components again.

The component_add_and_remove test tests how a server can add a component to an entity in the past, in this case a Shield, which prevents our enemy from taking damage.

TODO describe the various other tests.

Quick explanation of entity death in our rollback world

In order to preserve the Entity id, removing an entity using the DespawnMarker in fact removes all its registered components, but leaves the bare entity alive for rollback_frames.

Such entities retain their ComponentHistory buffers so they can be revived if needed because of a rollback. Finally, after rollback_frames has elapsed, they are despawn_recursived.

Removing components is hopefully a sufficient substitute for immediately despawning, however be aware the entity id will still exist until finally despawned.

Caveats:

• Developing this alongside a simple game, so this is based on what I need for my attempt at a server-authoritative multiplayer game.
• Currently requires you to use [GameClock] struct from this crate as frame counter.
• Littered with a variety of debug logging, set your log level accordingly
• Unoptimized: clones components each frame without checking if they've changed.
• Doesn't rollback resources or other things, just (registered) component data.
• Registered components must impl PartialEq
• I'm using a patched version of bevy_xpbd at the mo, to make Collider impl PartialEq (PRs sent..)