Chained consumer pattern

Instead of one mega-service calling LLM then DB, millipede uses chained consumers — each stage is a small deployable unit.

The chain

Click each link in ChainedConsumerDiagram:

HTTP → ingestion → raw-dev-events
       → llm-worker → enriched-dev-events
       → analyzer → Postgres + Redis
       → radar (poll + SSE)

Benefits

Benefit	How
Scale independently	Add llm-worker replicas
Swap enrichment	Change Python without touching Rust ingestion
Replay	Rewind consumer offset on enriched topic

Teach-back prompt

Could analyzer subscribe to raw-dev-events instead? When would you?