Struct Engine 

pub struct Engine {
    pub router: Arc<PartitionRouter>,
    /* private fields */
}

The FlowFabric engine: partition routing + background scanners.

Fields

router: Arc<PartitionRouter>

Implementations

impl Engine

pub fn start(config: EngineConfig, backend: Arc<dyn EngineBackend>) -> Self

Start the engine with the given config and backend.

Spawns background scanner tasks. Returns immediately.

backend is an Arc<dyn EngineBackend>. When the underlying concrete type is ff_backend_valkey::ValkeyBackend, the engine spawns its in-tree Valkey scanner supervisors. For non-Valkey backends (Postgres, SQLite, custom impls), the reconciler supervisor lives inside the backend itself (spawned during connect_with_*) and this constructor only wires the completion dispatch loop + partition router. Cairn #436: the runtime downcast-or-panic from v0.12 PR-7a is gone.

pub fn start_with_metrics( config: EngineConfig, backend: Arc<dyn EngineBackend>, metrics: Arc<Metrics>, ) -> Self

PR-94: start the engine with a shared observability registry.

Used by ff-server so scanner cycle metrics funnel into the same Prometheus registry exposed at /metrics. Under the observability feature (flipped via the same feature on ff-server / ff-engine), the handle records into an OTEL MeterProvider; otherwise the shim no-ops.

pub fn start_with_completions( config: EngineConfig, backend: Arc<dyn EngineBackend>, metrics: Arc<Metrics>, completions: CompletionStream, ) -> Self

Start the engine with a shared observability registry and a completion stream for push-based DAG promotion (issue #90).

The stream is typically produced by ff_core::completion_backend::CompletionBackend::subscribe_completions. The engine spawns a dispatch loop that drains the stream and fires ff_resolve_dependency per completion, reducing DAG latency from interval × levels to ~RTT × levels. The dependency_reconciler scanner remains as a safety net for completions missed during subscriber reconnect windows.

Backend parameter

backend is an Arc<dyn EngineBackend>. Works uniformly across Valkey / Postgres / SQLite (and any other impl): the engine attempts a Valkey downcast and, when successful, spawns the in-tree ferriskey-speaking scanner supervisors. Otherwise it trusts the backend to own its reconciler supervisor (cairn #436 / PR-7b). No runtime panic path.

Example

// Any `EngineBackend` impl works (cairn #436).
let engine = ff_engine::Engine::start_with_completions(
    cfg, backend, metrics, stream,
);

pub async fn shutdown(self)

Signal all scanners to stop and wait for them to finish.

Waits up to 15 seconds for scanners to drain. If any scanner is blocked on a hung Valkey command, the timeout prevents shutdown from hanging indefinitely (Kubernetes SIGKILL safety).