TL;DR
Get tech for your team delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Turbopuffer says it is redesigning its storage architecture for a planned v3, moving away from a layout organized around approximate nearest-neighbor vector search. The company says the change is intended to speed up text, regex and vector search and support more SQL-style queries; performance results and a release date have not been provided in the supplied report.
Turbopuffer says it is redesigning the storage architecture behind its database, moving its approximate nearest-neighbor (ANN) vector index from the primary position to a secondary index as part of a planned v3. Engineer Dan Harrison said the change is intended to improve text, regex and vector search while allowing more SQL queries to run efficiently on the service.
In a report published September 30, 2026, Harrison said turbopuffer is changing how documents and indexes are laid out, written, compacted and queried. The company described the work as underway, but did not give a release date or publish benchmark results for v3. The stated performance gains are goals, not reported outcomes.
Turbopuffer began as a serverless vector database, with a storage design built around ANN search. Its original approach used object storage as the source of truth and tiered NVMe solid-state drive and memory caches for performance. In later versions, the service added attribute filtering and full-text search, followed by capabilities including aggregations, regex search, fuzzy matching, sparse vector search and attribute ordering.
Harrison said the ANN-centered layout has become a constraint as customers use the product for query types beyond vector search. Under the proposed design, a new index would become primary and ANN would be treated as one of several secondary indexes. The post does not spell out the new index’s implementation or say how existing customer data and workloads will be migrated.
A Broader Role for Turbopuffer
The redesign reflects a shift in what turbopuffer is being asked to do: the company says customers now use it for text and regex search and other workloads, not just nearest-neighbor retrieval. If the new storage layout works as intended, it could make the service more suitable for applications that mix search, filtering and SQL-style operations, rather than requiring a separate system for each type of query.
The tradeoff is that the current architecture has been tuned for a core task that remains important to the product. Harrison said it supports indexes exceeding 100 billion vectors, with 200-millisecond p99 reads at more than 1,000 queries per second. Those figures are claims in the company’s report, not independently verified measurements supplied with the announcement. Reorganizing storage could improve other query patterns, but the company acknowledges that significant changes could cause regressions in ANN performance.
For customers, the practical question is whether v3 can broaden the database’s capabilities without degrading the vector-search performance and cost profile that helped attract users. The announcement describes the engineering rationale, but offers no comparative tests, pricing changes or customer migration details with which to judge the results.
vector database for machine learning
As an affiliate, we earn on qualifying purchases.
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, arranging vectors into clusters for search. The company chose a hierarchical clustering approach suited to object storage, initially using SPANN and later moving to SPFresh to support incremental indexing, according to Harrison’s account.
As customers asked for more features, the system added attribute indexes for filtering and postings-based indexes for full-text search. Those structures and later query capabilities were built around the existing ANN-primary storage layout. Harrison said that design remained largely unchanged even as the query engine expanded, leaving other operations constrained by choices originally made to serve vector search.
The announcement is therefore an architectural update rather than a report that vector databases as a category are obsolete. Its title, “RIP, vector database,” refers to turbopuffer’s planned change in how its own system is organized; the company says vector search will remain supported, but as a secondary index.
“We are changing turbopuffer’s storage architecture to take search to the next level.”
— Dan Harrison, turbopuffer engineer
As an affiliate, we earn on qualifying purchases.
Performance and Migration Questions
The report does not provide a v3 launch date, benchmark comparisons or measured performance from the new architecture. It also does not explain the mechanics of the new primary index, how customer data will be migrated, whether customers will need to change queries, or whether pricing and service limits will change.
It remains unknown whether gains in text, regex and SQL-style workloads will come with any measurable tradeoff in ANN speed, recall, reliability or cost. Harrison’s report says the redesign is intended to improve searches “in every respect,” while also warning that changes may regress ANN performance. Those statements describe the company’s aims and risks, not settled results.
As an affiliate, we earn on qualifying purchases.
V3 Design and Test Results
Turbopuffer says it is in the process of making the architectural change and plans to share updates on the work. The next useful details for customers will be an explanation of the new primary index, performance tests across the supported query types, and guidance on compatibility and migration.
Until the company publishes those details and results, the announcement establishes the direction of the redesign but not whether it will deliver the promised improvements. Customers evaluating the service will need to wait for information about availability and how v3 performs on their workloads.
As an affiliate, we earn on qualifying purchases.
Key Questions
Is turbopuffer ending vector search?
No. The report says the ANN index will become a secondary index, not that vector search will be removed. The redesign changes its role in the storage architecture.
What is turbopuffer changing in v3?
The company says it is changing how documents and indexes are organized, written, compacted and queried. It plans to replace the ANN index as the primary index with a new primary structure, which the report does not describe in detail.
What improvements does turbopuffer expect?
Engineer Dan Harrison said the work is intended to speed up text, regex and vector search and make more SQL queries run efficiently. The report gives no v3 benchmark results, so these remain goals rather than demonstrated gains.
When will turbopuffer v3 be available?
The September 30, 2026 report says the redesign is underway but gives no release date. It also does not specify a customer migration schedule.
Why could the redesign affect vector-search performance?
Turbopuffer’s current storage layout was built around ANN search and has been tuned for that workload. Harrison said a significant architectural change could cause regressions in ANN performance, although the report does not say that any regression has occurred.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
