Harvest, process, store, serve and present are separated so a connector failure degrades freshness for one source rather than availability for everyone.
The repository is deliberately four stores, not one: no single engine does versioned writes, graph traversal, ranked search and cheap archival well (view 08).
GraphQL is the primary read interface because the model is a graph; REST carries writes and bulk operations where GraphQL adds nothing.
Targets
5 million assets, 50 million relationships at design scale.
Search p95 under 300 ms; asset profile p95 under 500 ms; 3-hop lineage p95 under 1.5 s.
99.9% availability on the read path; ingest is allowed to lag at 99.5%.
Left out here
Identity, workflow, audit, notification and observability cut across every stage; drawing them here would obscure the flow. They are in views 03, 04, 13, 18 and 21.
