In this pipeline, aren't alternative approaches like turbopuffer efficient?

Meaning no much pruning in the space, but utilizing Object storage for long term retention, and cache-hit staying in nvme?

Curious how it compares as companies like Linear and Cursor are using it but not sure if it would help in this case.

Yes! You are right that with turbopuffer you eliminate the need for manual compression and pruning and you get ultra-fast response times out of the box. I tried to stay at a higher system-design level and I think the optimal choice depends heavily on an organization’s scale and maturity: startups vs large enterprises.