Braintrust Ditches Off-the-Shelf Stack, Builds Custom Database to Handle 100,000 AI Spans Per Second

Braintrust scraps its fragile three-part database stack and launches a custom-built solution called Brainstore, capable of ingesting 100,000 AI spans per second while delivering real-time, interactive query results at massive scale.

• Braintrust builds a custom database called Brainstore to handle the extreme demands of AI observability, including ingesting up to 100,000 spans per second, processing payloads up to tens of GB, and supporting real-time updates to long-running agent traces.
• A previous three-part stack using an open-source warehouse, Postgres, and DuckDB proves too fragile, suffering from multi-minute data latency, write bottlenecks, memory spikes, and correctness issues that worsen as scale increases.
• Brainstore solves these challenges through a write-ahead log for high-throughput ingest, asynchronous background compaction into multi-format indexes, per-customer data partitioning on object storage, and a query pipeline that merges indexed and unindexed data to deliver real-time, interactive results.