VCI looks promising for the "don't ship your data to ClickHouse just to run aggregates" case, and an in-memory columnar index covers a real slice of it. Where I'd temper expectations is that a columnar index bolted onto a row-store engine still inherits that engine's execution model and storage, so you usually land at "good enough for dashboards on recent data" rather than ClickHouse-class scans over cold history. The columnar layout was never really the hard part. It's the storage tier and the query engine underneath it.

A little shameless self-promotion, but this gap is exactly what I've been building FabricDB around. It's built on top of FlightSQL and leverages postgres under the hood (could also implement postgres wire protocol some day, but that would eat some of the OLAP benefits you get with DataFusion, arrow, ADBC, etc. that make fabric appealing), keeps the analytical side in object storage you own, and serves both transactional and analytical queries from one system, so there's no second engine and no CDC. Disclosure, I'm the founder here, so take it with a grain of salt. There's an early whitepaper if anyone wants to pick at the architecture: https://fabricdb.io/whitepaper.html