Perplexity introduced CobbleDB, a custom key-value database designed for AI-powered search, on September 14, 2026. The system, written in approximately 40,000 lines of Rust, replaces the previous architecture based on Amazon DynamoDB.
The transition resulted in significant performance gains. Median batch-read latency decreased from 31.4 ms to 5.60 ms, marking an 82% improvement. At the 90th percentile, latency dropped from 56.7 ms to 9.77 ms, while 99th-percentile latency fell from 123 ms to 24.2 ms.
The database architecture consists of three components: Pillar for document durability, Lorry for batched ingestion, and the serving layer for low-latency reads. The system utilizes RocksDB for transactions, incorporating custom partitioning and caching strategies.
Perplexity utilized internal AI coding agents to develop the database, completing the project in approximately two months. Internal projections indicate that CobbleDB could lower storage-layer costs by over 20% compared to DynamoDB by removing per-byte pricing.
The company plans to open-source CobbleDB once the system proves its stability in production environments handling hundreds of thousands of requests per second.
