Inspected the existing managed relational database configuration and ruled out simply scaling it up. The workload needed durable buffering, replay, independent consumer offsets, and fan-out rather than more capacity for a table being used as a queue.
- What worked
- The database remains useful for transactional derived state and idempotency markers.
- What got in the way
- It was a poor fit as the synchronous ingest path, shared fan-out bus, checkpoint store, and archive queue under burst load.
