turbopuffer
Search and retrieval with object storage as the sole source of truth.
turbopuffer is a search engine that treats object storage as the primary store rather than a backup. Vectors, full-text indexes, and documents live in S3-compatible object storage, and a cache tier on NVMe SSDs serves hot queries with low latency. Storage and compute are cleanly separated, which is exactly the lakebase pattern applied to retrieval instead of transactions.
Why it qualifies
- Object storage is the source of truth. Data durability lives in object storage (S3, GCS, or Cloudflare R2), not on disks tied to a specific node.
- Storage and compute are separate. A caching layer holds working data close to compute for speed, but the durable record sits in object storage, so compute can be added, removed, or replaced freely.
- Operational workload. turbopuffer serves live, low-latency queries for applications and AI systems, not batch analytics.
Where it sits in the category
turbopuffer is a search and retrieval system, not a transactional (OLTP) database. It fits the lakebase category under the broader reading used on this site: an operational database built on object storage that serves live application and agent queries. The economics match the relational case. Because the durable data lives in cheap object storage, keeping a large index around costs far less than pinning it all to memory or local disk.