Turbopuffer宣布重构其存储架构并推出v3版本 1。新架构的核心变更是停止以向量近似最近邻（ANN）索引作为主键，转而采用通用的键值存储方式，使向量索引成为辅助性索引 1。
这一架构调整旨在解决三个关键问题 1：存储放大、写入放大以及向量化效率受限。Turbopuffer在全文搜索（FTS）v2上的优化中实现了索引大小减少10倍、查询速度提升20倍的效果 1。Turbopuffer v3已通过全部持续集成（CI）测试，但目前处于性能优化的早期阶段，存在显著的性能回退 1。
截至公告时，Turbopuffer的规模已达到托管1T以上文档、每秒处理1000万次以上写入操作、每秒处理25000次以上查询 1。
Turbopuffer has announced a fundamental redesign of its storage architecture with the release of version 3, moving away from vector indexing as its primary organizational principle 1. The new system implements a general-purpose key-value storage model, relegating approximate nearest neighbor (ANN) indexing to a secondary role alongside other index types 1. This architectural shift aims to address three critical limitations of the previous design: storage amplification, write amplification, and inefficient CPU vectorization 1.
The restructuring enables Turbopuffer to support a broader range of query types and handle significantly larger datasets 1. The company's v3 version has achieved 100% continuous integration test passage earlier this month 1, though the platform is currently in performance optimization phases with documented regression in initial benchmarks 1. At its current scale, Turbopuffer manages over 1 trillion documents while processing more than 10 million writes and 25,000 queries per second 1. Previous optimizations to full-text search functionality demonstrated the potential of this approach, reducing index size by 10 times and improving query speed by 20 times 1.
评论
还没有评论，欢迎留下第一条。