executionruntime

Execution-aware JobDB runtime

Wrap a persistent JobDB runtime with executionruntime.New(runtime, allocation, onHandoff). Pass the same normalized allocation and callback to compiler.RecipeJobWorkerOptions, with StageExecution: wrapped.Stage and WrapTaskWorker: wrapped.WrapTaskWorker, and build the engine using the wrapped runtime. Register schemas using the original runtime. The CLI performs this wiring for all execution entry points.

The adapter checks current published demand at lease admission, including task-only leases. An incompatible lease is rescheduled with the same route and task coordinates, without rewriting the snapshot. Recipe preflight separately resolves and checks the pinned base when no snapshot has been published.

For recipes using execution_needs, recipe leases replay first: an older published scope may belong to completed work. The compiler stages the inherited needs before each task call, and guarded task workers check them only after JobDB has decided the task is unfinished. Completed results bypass the worker and do not request historical environments. This also applies when recovery starts through a task route: its saved task may already have completed before the executor crashed. The next genuinely unfinished task is checked instead.

If building a workset manually, wrap every task worker with WrapTaskWorker and supply that same function in the recipe worker options. This is required for scoped needs; staging alone cannot enforce a live-task boundary.

Node scopes are ephemeral. Any subsequent ordinary reschedule publishes the staged scope if it differs from the saved snapshot, without mutating permanent job overrides. Explicit job requirements still override recipe and node declarations. No live scheduler updates or automatic downsizing are performed.

The compiler stages resolved demand so ordinary task/time reschedules can publish it if needed. An explicit requirement change is a full revision-checked snapshot on yield. The adapter preserves lease token/worker identity forwarding needed for chapter persistence; it does not add lease matching or provisioning.

Stop or redirect an incompatible worker after OnHandoff to avoid a tight reacquisition loop. The event is a handoff, not completion. Admission errors stop keepalive for rejected leases; ordinary lease-expiry recovery applies.

Use ops.SuspendExecution at a successful durable activity boundary to request a change. Read-only replay must use ReadOnlyReplay: true and must not publish. See the design and user guide for lifecycle and failure semantics.