Chaos Search seems to be doing this architecture already and according to the podcast episode [1], it uses a highly optimized storage layout.

Never used it, so would be interested if somebody could comment on it.

[1] https://www.dataengineeringpodcast.com/chaos-search-with-pet...