Dan Harrison (Engineer) announced that turbopuffer is changing its storage architecture with turbopuffer v3. The new engine changes how documents and indexes are laid out, written, compacted, and queried, allowing many more query plans at much greater scale. turbopuffer launched as a serverless vector database, with object storage as the source of truth for economics and tiered NVMe SSD/memory caches for performance. This was validated by early customers including Cursor and Notion.
Over time, turbopuffer evolved into a generalized database for search, supporting attribute filtering and full-text search. The query engine evolved, but the storage architecture remained centered on the ANN index as the primary index. Now, they are moving to a new primary index, making ANN "just another" secondary index.
In v1, documents were just an ID and a vector, using a hierarchical clustering index. v2 added attribute filtering via inverted indexes and full-text search. The team will share updates as they transition. "We've pushed the primary ANN index as far as we can, but it's time to move on."
Source: Hacker News · Summarized by HeadlinesBriefing