jandrewrogers' point about cache replacement pathology in graph traversals is the kind of thing you only learn after burning real hardware cycles. I hit similar issues using Neo4j for knowledge graph queries — once traversal depth exceeded 4-5 hops, latency became unpredictable regardless of data size.

The embeddable angle is interesting though. For small-to-medium graphs (< 10M edges) that live in a single process, these scalability concerns matter less, and the developer experience of "just add a dependency" beats running a separate database server.