Repository navigation
Spawned 0.8 adds deferred replies and actor-owned jobs, and aligns exit signals, supervision startup and runtime-drop cleanup with the documented OTP lifecycle contract. This is a breaking release; see the 0.7 → 0.8 migration guide.
Main changes
- Typed deferred replies through
PendingReply,ReplyToandPendingReplieslet an actor keep handling messages while work finishes elsewhere. - Pipe-to-self and reply jobs support keyed replacement, deadlines, bounded outstanding work and lifecycle fencing.
- One death point answers outstanding callers and fences signals before post-death payload destructors, wakers or tracing callbacks run.
- OTP-style exit signals include direct
kill, trappable linkedKilled,NoProc, fenced unlink/demonitor and explicit stopped-hook behavior. - Supervisor restart budgets count recovery rounds; failed initial child startup rolls back the tree.
- Deterministic simulation adds clock/entropy interception in a separate simulation-only crate, inline blocking work and runtime timer/JoinSet adapters. Platform limits and migration requirements are documented.
Upgrade
Handler::handle now receives a reply token; send-only handlers use #[send_handler]. Raw requests use send_request with Deadline; use Deadline::Never for the former unbounded raw-request behavior. Ordinary request() retains its five-second timeout. Only caller expiry decides timeout, so a late reply may still win before expiry is processed.
The release also contains fixes for stop/receive races, deferred-job payload destruction, simulated deadline ownership, deadline-worker initialization and allocator-independent link-lock ordering. Core crates retain #![forbid(unsafe_code)]; no coarse clock is introduced.
See the changelog, examples and verification limits.
Scheduling limitation
Managed and unmanaged actors share a thread for heap-driven deadlines. A blocking user destructor or waker there can delay other actors' expiry and scheduled job emits. Ordinary request/Response timers and send_after/send_interval drivers are separate. The guide states the callback-progress requirement; callback isolation is tracked as follow-up work. Death-point caller-before-queued-payload/state ordering is unchanged.
Validation
All 21 required checks and both final ethrex/ethlambda gates passed, including default-stack debug RPC tests. The unchanged performance gates passed against v0.7.3, and fresh paired four-hour soaks passed with 43.4 million candidate actor starts, 14.4 million scenarios and zero workload errors. Independent Claude reviews, production-core Loom, simulation and fresh lifecycle trace checks passed. Attached evidence retains all original samples, the single fixed precision refinement, bounded-model provenance and the final release-tree mapping.