LangSmith We built SmithDB, the data layer for agent observability
Migrate to SmithDB for faster trace queries and full‑text search.
Switch your LangSmith ingestion to SmithDB and update queries.
Summary
SmithDB, launched on 13 May 2026, is LangSmith’s purpose‑built distributed database for agent observability and evaluation workloads. Built in Rust with Apache DataFusion and Vortex, it stores trace data in object storage, uses a stateless ingestion service, and a small Postgres metastore for metadata. Performance benchmarks show P50 latencies of 92 ms for trace tree loads, 71 ms for single run loads, 82 ms for run filtering, 630 ms for ingestion, and 400 ms for full‑text search, delivering 12–15× speedups over previous storage. SmithDB scales by adding compute rather than disks, making it easy to deploy in self‑hosted or multi‑cloud environments. It supports advanced query patterns such as random access, interactive filtering, full‑text search, JSON filtering, tree‑aware queries, thread reconstruction, and aggregations, all at low latency over large agent traces.
Key changes
- Rust implementation using DataFusion and Vortex for query execution
- Object‑storage backend with stateless ingestion and query services
- P50 latencies: 92 ms trace tree load, 71 ms single run load, 82 ms run filtering, 630 ms ingestion, 400 ms full‑text search
- 12–15× performance improvement over previous storage
- Scales by adding compute, no local disk management
- Supports random access, interactive filtering, full‑text search, JSON filtering, tree‑aware queries, thread reconstruction, and aggregations
- Fully self‑hosted and multi‑cloud compatible