Why turbopuffer v3 leaves vector-first storage
turbopuffer v3 is redesigning turbopuffer's storage architecture. The product began as a serverless vector database built around object storage, using tiered caches to control cost and an approximate-nearest-neighbor (ANN) index to organize documents, attributes, and query plans. As it added attribute filters, BM25 full-text search, regex, fuzzy matching, sparse vectors, and aggregations, that design increasingly became a constraint.
From ANN primary index to a broader query foundation
The old layout stored a document's full contents under its ANN address. Multi-vector documents could therefore duplicate data, while vector rebalancing could move attributes and full-text indexes along with the document, increasing write amplification. Query plans were also forced into blocks sized for vector clusters, limiting batching, compression, and SIMD use. The article says that separating postings into fixed-size blocks previously made the full-text index ten times smaller and queries up to twenty times faster.
The central change in v3 is to replace the primary index, making ANN just another secondary index. This should let text, attribute, aggregation, and scan workloads use layouts and block sizes suited to their own access patterns, while creating a foundation for faster SQL queries. The team says all continuous-integration tests now pass and that it will publish performance work as it seeks parity with, and eventually improvement over, the current system.
The key question is vector-search performance
This is not a cost-free abstraction change. The existing vector-first architecture has supported very large ANN workloads, so a storage rewrite could affect recall, latency, or throughput. The public update establishes a correctness milestone, but not yet complete production benchmarks, migration details, or regression data. The important question is therefore not whether vector databases are disappearing, but whether the new primary index can preserve ANN performance while materially improving hybrid search and non-vector queries.