Knowledge graphs, GraphRAG, and LangGraph solved three problems. The fourth is the work itself: a reviewable graph of bounded agentic loops.
Trace what one tiny Iceberg commit writes, then model hourly, per-minute, and per-second cadences so streaming costs become arithmetic, not adjectives.
Position deletes, equality deletes, and deletion vectors compared from the Iceberg spec: what each writes, how readers apply it, and when to use which.
Iceberg is turning from a JVM table format into a library other systems embed. What that shift changes for engines, catalogs, and the spec itself.
