Copy-on-write versus merge-on-read in Iceberg, how v3 deletion vectors track row-level deletes, what readers pay, and compaction patterns for high-churn tables.
Why JSON strings and map columns fall short for LLM outputs and agent logs, and how a native Iceberg variant type stores and queries nested AI payloads instead.
How Iceberg optimistic concurrency behaves when fleets of agents commit, and how partition isolation, commit queues, and ingestion gateways prevent conflicts.
Why stateful MCP servers fail to scale under agent load, and how stateless gateways handle credential delegation, context isolation, and database guardrails.
