turbopuffer explains why it is changing its primary index

turbopuffer is moving toward a storage design where approximate vector search is a secondary index instead of organizing every document. Today, changing vector clusters can move document attributes and related indexes too. The same layout also duplicates content for multiple vectors and restricts query batch sizes. The planned redesign aims to remove those constraints for broader search and aggregation workloads.

turbopuffer · Sep 30, 2026

Why it matters

The article shows how an early indexing choice can constrain later features and operating costs.

More context

Read the engineering walkthrough for its storage and write-amplification examples. Watch the v3 series for implementation and migration details; this first post describes an architectural direction rather than a generally available replacement.

Keep in mind

Vector search is not being removed despite the headline. Expected improvements are the vendor’s goals, and changing the layout risks regressions. 1min has not benchmarked either architecture.

Original read