TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Turbopuffer has announced a v3 storage redesign that will make its approximate-nearest-neighbor vector index secondary to a new primary index. The company says the change is intended to improve text, regex and vector search and support more SQL queries; performance results and a rollout schedule have not yet been provided.
Turbopuffer is redesigning its storage architecture for v3, moving away from a layout built around its approximate-nearest-neighbor (ANN) vector index. In a September 30, 2026 company post, engineer Dan Harrison said vectors will become a secondary index under a new primary index, a change intended to improve several kinds of search and make more SQL queries fast.
The change addresses a mismatch between turbopuffer’s original design and how customers now use it. The service launched as a serverless vector database, with storage organized to support vector search on object storage. The company later added attribute filtering, full-text search, regex, fuzzy matching, sparse-vector search, aggregations and attribute ordering, while retaining the original vector-centered storage layout.
Harrison said the new v3 architecture changes how documents and indexes are laid out, written, compacted and queried. The company’s stated goals include faster text, regex and vector search, as well as support for more SQL queries. The post identifies the existing design as a constraint on query plans such as GROUP BY and aggregations.
The announcement describes work in progress, not a completed migration or a measured performance release. Turbopuffer did not provide benchmark results for v3, a release date, or details of how existing users will move to the new layout in the material available here.
A Broader Role Beyond Vector Search
The redesign reflects a shift in the service’s intended role: from a database whose storage is centered on vector retrieval to a platform that must handle multiple query types. That matters to customers using it for text search, filtering, analytics-like queries or other workloads, because the underlying data layout can affect how efficiently those operations run.
For existing vector-search users, the change also carries a trade-off. Turbopuffer says its current architecture performs well for ANN workloads, and Harrison wrote that significant changes could introduce regressions in vector performance. The company will need to demonstrate that a broader indexing model improves non-vector workloads without undermining the performance customers rely on.
The announcement is relevant to database teams evaluating whether one system can serve search and additional data-query needs. It is not yet evidence that v3 will meet those goals: the company has stated its rationale and direction, but published no comparative results in this report.
As an affiliate, we earn on qualifying purchases.
From Vector-First to Multi-Query
Turbopuffer’s first version stored documents as an ID and a vector. Its vector data was arranged in clusters, with a hierarchical clustering index designed to work with object storage. The company initially used SPANN and later moved to SPFresh to support incremental indexing. In that design, a document’s cluster and local identifiers formed an ANN address, which served as the organizing key for the storage layout.
In the service’s second phase, attribute filtering and BM25 full-text search added new query plans. Attribute indexes mapped values to document addresses, while text indexes stored postings and scoring information. Other capabilities followed, but, Harrison said, they were built around the same vector-primary architecture. The September announcement marks the planned departure from that arrangement.
Harrison said turbopuffer has served vector indexes containing more than 100 billion vectors, with 200-millisecond p99 reads at more than 1,000 queries per second. Those figures are company-reported performance claims, not independently verified results in the source material. He cited early customers including Cursor and Notion, and said Linear uses the service for a syncing engine.
“We’re in the process of moving to a new primary index, and making ANN “just another” secondary index.”
— Dan Harrison, turbopuffer engineer
Performance and Migration Details
The announcement does not specify when v3 will ship, when customers can test it, or whether the transition will require changes to applications or data. It also does not describe the new primary index in detail or provide measurements comparing v3 with the existing system.
It remains unclear how the redesign will affect the service’s established ANN workloads, including latency, throughput, cost and recall. Turbopuffer acknowledges that changing the architecture could harm vector performance, but gives no migration safeguards or results in this update. Its stated improvements are goals, not demonstrated outcomes so far.
Look for Benchmarks and Rollout Plans
Turbopuffer said it is in the process of changing the storage architecture and plans to share updates on the work. The next useful milestones for customers will be technical details about the new primary index, test or release availability, and benchmarks across vector, text, regex and SQL workloads.
Customers considering the change will also need information about compatibility, migration steps, pricing or resource effects, and how the company will track regressions. Until turbopuffer provides those details and performance data, the practical impact of v3 remains unconfirmed.
Key Questions
What did turbopuffer announce?
It announced a v3 storage redesign that replaces its vector-centered primary index with a new primary index and makes the ANN vector index secondary.
Why is turbopuffer changing its architecture?
Engineer Dan Harrison said the existing layout works well for vector search but constrains other query plans, including GROUP BY and aggregations. The redesign is intended to support a broader mix of search and SQL workloads.
Will v3 make vector search faster?
Turbopuffer says faster vector search is a goal, along with faster text and regex search. The announcement includes no v3 benchmark results, so the performance improvement has not yet been demonstrated.
When will customers get v3?
The company said the work is in progress but did not give a release date or customer migration schedule in the announcement.
Does this mean turbopuffer is no longer a vector database?
No such product-status change was announced. The company plans to make its ANN index secondary in the storage architecture, while continuing to cite vector search as one of the search workloads v3 is meant to support.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
