turbopuffer v3 Overhauls Storage Architecture, Decoupling Vector Indexes

turbopuffer v3 Storage Redesign

turbopuffer is implementing a new storage architecture in its upcoming version 3. This redesign changes how documents and indexes are laid out, written, compacted, and queried within the system. The primary goal is to enhance search performance across the board, including vector search capabilities.

Enabling Broader SQL Query Support

Beyond improving search, the new architecture also establishes a foundation for executing a wider range of SQL queries more quickly within turbopuffer. This expansion of functionality suggests a move towards supporting more diverse data operations than previously possible.

Evolution from Vector-Primary Design

Initially launched as a serverless vector database (v1), turbopuffer specialized in cost-effective and fast vector searches, utilizing object storage and tiered caching. Version 2 added strong text and regex search. However, the original storage architecture, where the ANN vector index was the primary index, constrained certain query plans like GROUP BY and aggregations. The v3 update addresses these limitations by making the ANN index a secondary index.

Historical Indexing Approach

In turbopuffer v1, documents consisted of an ID and a vector. The system used a hierarchical clustering index, starting with SPANN and later migrating to SPFresh for incremental indexing. This method clustered vectors into groups, forming a tree structure with centroids.