turbopuffer

Search infrastructure company building turbopuffer Search DB, a search database whose index lives in object storage with a local cache hierarchy over it. Positioning is cost and scale at large corpus sizes — see Compute-Storage Disaggregation for why that architecture makes a different set of trades than Elasticsearch-style shared-nothing clusters or the local-disk vector databases (Qdrant, Weaviate, Pinecone).

Notable Contributions

Single-WAL multi-shard design (2026) — sharding for an object-storage-native search engine that avoids distributed commit entirely. One write-ahead log per namespace gives a single global sequence order for writes; each of N shards consumes 1/Nth of that log, so indexing and queries scale out without coordination between shards. Two-phase commit was rejected on the stated grounds that its round-trip cost is amplified by object storage latency, and that “simplicity scales.” Raised the per-namespace ceiling from 1 TB (SaaS) / 4 TB (custom) to 256 TB. See How to Build a 256 TB Search Index.

Ephemeral-NVMe cache tiering — the argument that rebalancing cost is a property of stateful nodes rather than of large shards. If local state is a disposable cache refillable at 2–4 GB/s from S3, large shards stop being operationally painful, which is what makes the 256 TB ceiling practical.

People

Articles