Turbopuffer

Turbopuffer is a serverless vector and full-text search engine built on object storage (S3/GCS), founded by ex-Shopify engineer Simon Eskildsen. Its core architectural bet is to keep the system of record on cheap object storage (~$20/TB/month) while promoting hot data to NVMe SSD and RAM caches based on access patterns. This three-tier hierarchy lets it claim in-memory-class latency on warm queries (~10ms p90 for vector, ~18ms p90 for BM25 on 1M docs) at roughly 10x lower cost than RAM-resident incumbents like Pinecone. The tradeoff is high cold-query latency (~444ms p90) on first access and relatively high write latency (~248ms p90 for 512KB upserts), since durability is anchored to object storage. It is production-proven at notable scale, powering Cursor, Notion, Linear, Ramp, and others, and the company reports handling 4T+ documents, 10M+ writes/s, and 25k+ queries/s across its fleet.

The ideal user is a team running large-scale RAG, semantic search, or hybrid (vector + BM25 + filter) retrieval where storage volume dominates cost and where occasional cold-start latency on rarely-touched namespaces is acceptable. Multi-tenancy is a first-class concern: namespaces are cheap and isolate per-tenant data, making it a strong fit for AI products that shard search by customer. It is a poor fit for ultra-low-latency, always-hot workloads that need single-digit-ms tail latency on every request regardless of access pattern, or for teams wanting a generous free tier, there is none, and the floor is a $64/month minimum (Launch tier minimum was reduced to $16/month in a later update). Critics (notably Zilliz) argue the object-storage tradeoff hides latency and operational consequences that aren't captured by storage-cost-per-GB alone.

The community sentiment skews positive on operational simplicity, recall auto-tuning, and the engineer-grade product, with the common heuristic being "use pgvector until it breaks, then go to Turbopuffer." The main reservations are pricing for cost-sensitive users versus pgvector, the absence of a free tier, and the inherent cold-query penalty. The product remained invite/waitlist-gated for much of its early life before broader GA, which slowed evaluation for some teams.