Event-Streaming Backbone
An append-only, partitioned event log is the single source of truth carrying every measurement, prediction, proposal, and command through the stack.
Why a log, not a bus
Orchestration is built on an append-only, partitioned commit log rather than a fire-and-forget message bus. Every fact about the breeder (Hyperion) and burner (Aegis / MetroVolt) enters as an immutable, sequenced event. Consumers read at their own offset, replay history deterministically, and rebuild state from scratch. This is the substrate that makes deterministic replay and decision lineage possible.
Topic and partition model
Events are grouped into topics by domain and partitioned by a key that preserves the ordering that physics requires. For the breeder, magnetics telemetry partitions by coil-set so that a single coil's current history is a totally ordered sub-stream; for the burner, plug and throat diagnostics partition by mirror-cell. Partition key selection is a correctness decision, covered in event ordering and partitioning.
topic: breeder.magnetics.coil_current key=coil_id
topic: breeder.plasma.equilibrium key=shot_id
topic: burner.plug.field key=cell_id
topic: *.commands key=actuator_id (single-writer)
topic: *.proposals key=shot_id
topic: *.audit append-only, never compacted
Retention and compaction
- Raw diagnostic topics: time-retained, then tiered to cold storage for post-campaign analysis.
- State topics (equilibrium reconstruction, twin snapshots): log-compacted so the latest value per key is always recoverable.
- Command and audit topics: infinite retention, cryptographically chained, never compacted.
Throughput budget
The backbone is sized for the burst profile of a shot, not the average. A breeder disruption event can emit tens of thousands of high-rate magnetics samples in milliseconds; the log absorbs the burst and lets slower consumers fall behind gracefully under backpressure rather than dropping data. Real-time control on L1 never depends on the log for its microsecond loop; the log is the coordination and record plane, not the deterministic control path.
Producers are idempotent (see idempotency) so that a retried write after a network partition does not duplicate a measurement or, worse, a command.