TurboPuffer, a serverless vector database platform, is undergoing a fundamental architectural redesign that eliminates the vector index as the primary organizational structure for stored data.
According to the company’s engineering blog, the shift marks the end of what it calls the “vector-primary architecture” that has defined the platform since its 2023 launch. In v1, TurboPuffer specialized exclusively in cheap, fast vector similarity search using hierarchical clustering indexes (SPFresh) layered on object storage. The platform expanded in v2 to support attribute filtering, full-text search, regex matching, and other query types—but all of these remained built around the same vector-primary storage layout.
“We’ve pushed the vector-primary architecture as far as we can, and it’s time to move on,” the company states. The new v3 architecture makes the approximate nearest neighbor (ANN) vector index “just another secondary index” rather than the primary one, allowing the storage engine to optimize for multiple query types simultaneously.
Three constraints drove the redesign. First, storing full document contents under ANN addresses causes storage duplication when documents have multiple vector representations. Second, when vectors are rebalanced to maintain clustering quality, the entire document and all its associated indexes must move—a cascading write amplification problem that has become a bottleneck. Third, block sizes across all query engines are constrained by ANN cluster sizes (100–200 documents), preventing modern vectorized execution patterns that work best with much larger blocks (DuckDB uses 2,048-row batches; ClickHouse up to ~65,000).
The company cites its own full-text search v2 as evidence of the block-size impact: reworking postings into fixed 256-document blocks instead of ANN-constrained 1.5-posting blocks reduced index size 10x and sped queries up to 20x.
TurboPuffer achieved a major milestone this month: 100% of its continuous integration test suite passes on v3. The company is now moving from correctness validation to performance optimization. The redesign is intended to accelerate search across text, regex, and vector modalities, while also making SQL queries faster and supporting additional SQL patterns like GROUP BY and aggregations that the vector-primary design constrained.
Key facts
- TurboPuffer is replacing its vector-primary storage architecture with a new design that treats ANN indexing as a secondary index
- The redesign aims to improve performance on text search, regex, and SQL queries, not just vector similarity
- Storage amplification, write amplification, and vectorization constraints drove the architectural change
- TurboPuffer v3 has achieved 100% CI test pass rate and is moving into performance optimization phase
- The platform will now support more efficient block sizes for different query types, enabling faster execution patterns