SmithDB, the data layer for agent observability
2 days ago
- #database
- #observability
- #agents
- SmithDB is a purpose-built distributed database for agent observability, designed to handle modern agent traces that are large, nested, and multi-modal, with complex query patterns.
- It delivers high performance with significant latency improvements, such as P50 latencies of 92ms for trace tree loads and 400ms for full-text search, making LangSmith experiences up to 15x faster.
- The architecture is portable and scalable, leveraging object storage with stateless services, making it easy to deploy in self-hosted and multi-cloud environments by adding compute rather than managing local disks.
- Key components include ingestion services, a metastore, query services, compaction services, and a cluster manager, optimized for agent observability workloads like progressive querying and late materialization of large fields.
- SmithDB now handles 100% of US Cloud ingestion and query traffic in LangSmith, with plans to fully integrate into all product surfaces and enable self-hosted deployments soon.