Nathan VanBenschoten

Chief architect at turbopuffer, working on scalable vector search. Previously a principal engineer at Cockroach Labs, working on CockroachDB’s transactions and replication — specifically the performance of its transaction, replication, and persistence layers.

That background is worth knowing when reading his search work: the design argument in How to Build a 256 TB Search Index is a distributed-transactions argument imported into search infrastructure. The decision to reject two-phase commit in favour of a single ordered write-ahead log consumed in parallel by N shards is the reasoning of someone who has paid 2PC’s coordination costs in a database and does not want to pay them again on top of object storage latency.

Articles in This Vault

Key Contributions

  • Single-WAL multi-shard architecture for object-storage-native search — total write ordering without distributed commit (Sharding, Compute-Storage Disaggregation)
  • The argument that ephemeral NVMe caching removes rebalancing as a constraint on shard size