Turbopuffer, a serverless vector database, has announced a significant change to its storage architecture with the upcoming "turbopuffer v3." The update shifts the core design from using Approximate Nearest Neighbor (ANN) as the primary index to treating it as just another secondary index.
Turbopuffer announces shift from ANN-primary to secondary indexing in v3 architecture
Since its launch, turbopuffer has utilized a hierarchical clustering index optimized for object storage, where documents are keyed by their ANN addresses. While this approach has enabled efficient vector searches at scale, the company states it has become a bottleneck for non-vector query types such as attribute filtering and full-text search.
The shift to a new architecture addresses three main limitations of the current system: storage amplification, particularly for multi-vector representations; write amplification caused by rebalancing clusters; and limited vectorization. Currently, updating a single vector can force the movement of large amounts of metadata and related indexes because all data is tied to the ANN address.
The new architecture aims to allow for more efficient block sizes tailored to different query plans, enabling better CPU saturation and SIMD utilization. Turbopuffer reported that while the v3 engine has passed CI tests, it is currently undergoing performance tuning to eliminate regressions compared to the existing production environment.
Sources
- RIP, vector database （Hacker News Frontpage, 2026-10-01）