Why We Didn't Converge: ClickHouse's VLDB Paper and the Architecture Agents Actually Need

ClickHouse’s new VLDB paper shows its performance comes from a stack of choices: MergeTree storage, vectorized execution, LLVM JIT, and broad file-format support.

The broader industry is shifting away from single all-in-one HTAP systems toward separate OLTP and OLAP architectures.

Snowflake and Databricks reflect the scale-first analytics path, while CockroachDB and Lakebase point to tightly integrated, low-latency consistency.

At the same time, AI agent workloads are pushing some vendors back toward more converged database designs.