turbopuffer rewrites storage engine v3 in public, grinding query plans live
vboykis · x · 2026-10-01
Vector database company turbopuffer announced a major overhaul of its storage architecture (v3) to serve more query plans at greater scale.
- As of 2026-09-05, v3 passes 100% of CI, but performance regressed significantly vs production (v2)
- The team is "grinding query plans in public": dev log updates live, with charts updating from day zero until parity
- Hot query regressions are large: hot full-text search p90 2264ms vs 181ms (12x slower), hybrid search 7x slower, attribute ordering 5-7x slower
- Hot vector search is at parity (1x); cold vector queries are slightly faster on v3 (0.83x)
A rare transparent, in-public view of a large-scale system rewrite.
More from Infra
- Case study: how Canva saved millions in cloud costs — _jaydeepkarale · 2026-10-01
- tilelang: A DSL for High-Performance GPU/Accelerator Kernels Hits 7,973 Stars — tile-ai · 2026-10-01
- Bain: AI companies need $4.2 trillion a year in new revenue by 2031 to fund data centers — ylecun · 2026-10-01
- Dev ditches AWS vector service for open-source Weaviate after scaling pain — CShorten30 · 2026-10-01
- Nebius acquires Inferize to cut the GPU idle tax in production inference — demian_ai · 2026-10-01
- One chart of the AI infra buildout: 2026 construction boom, 2030 bet on demand — AntDX316 · 2026-10-01