Trait CheckpointFlushObserver 

pub trait CheckpointFlushObserver {
    // Required methods
    fn on_flush_start(&self, lease_id: u128);
    fn on_flush_complete(&self, lease_id: u128);
}

Sideband observer notified at checkpoint flush start and complete.

Slice 96 wires this so the scheduler-service PreemptionManager can mark a lease as NonPreemptibleReason::CheckpointInProgress for the duration of a flush. The trait is intentionally minimal and observational — implementations MUST NOT block, return errors the flush can act on, or otherwise change the flush’s success semantics. A failed observer notification (e.g. lease not tracked by the preemption manager because it was released / expired / fenced concurrently) is a sideband condition the observer logs internally; it does NOT abort the checkpoint flush. The checkpoint is the source of truth for stage durability; the observer is a sideband signal to the scheduler.

Why a trait callback rather than a direct cross-crate call? grafos-pipeline is no_std + alloc and intentionally has no dependency on grafos-scheduler-service (which is std-only and depends on the http server, audit chain, etc.). A direct call would either force the pipeline crate into std or pull scheduler-service into no_std builds. The trait keeps both sides cycle-free: the producer (pipeline) calls a small trait; the consumer (scheduler-service) implements it.

Lifecycle contract:

• on_flush_start(lease_id) is called exactly ONCE at the beginning of a checkpoint() call when both the observer and lease_id are wired.
• on_flush_complete(lease_id) is called exactly ONCE at the end of checkpoint() — on BOTH the success path and every failure path. The implementation is responsible for the RAII guard via a Drop-based scope; producers don’t need to sprinkle clears at every error site.

Required Methods

fn on_flush_start(&self, lease_id: u128)

Called at the START of a checkpoint flush. Implementations should mark the lease as transiently non-preemptible.

fn on_flush_complete(&self, lease_id: u128)

Called at the END of a checkpoint flush (both success AND failure paths). Implementations should clear the non-preemptible marker.

Implementors