Perplexity swaps DynamoDB for in-house CobbleDB to speed up search reads
The Rust key-value store cuts batch-read latency in production and is slated for open-sourcing.
Perplexity has shifted core search reads from DynamoDB to CobbleDB, a Rust key-value store designed for batched retrieval of large language-model context; a query can involve 100–120 page keys, and records average about 50 KB. The architecture assigns durable data to Pillar, batch-file preparation to Lorry, and serving lookups to CobbleDB, using Amazon S3 to carry exported payloads. In production, median batch-read latency decreased from 31.4 ms to 5.60 ms, p90 from 56.7 ms to 9.77 ms, and p99 from 123 ms to 24.2 ms; storage costs fell by at least 20%. Tests with payloads up to 100 KB reached 500,000 requests per second, although asynchronous replica updates mean the system accepts some consistency lag. CEO Aravind Srinivas said two systems engineers and AI coding agents built the 40,000-line database in two months, and Perplexity plans to open-source it in a future release.
Why it matters
Moving core search reads onto CobbleDB means Perplexity can serve large language-model context faster by splitting durable storage, batch preparation and low-latency lookups into separate systems and tuning each for its role. The new store has cut batch-read latency in production and hit high request rates in tests, and Perplexity plans to release it as open source, making the design and implementation available for others to examine and reuse.
Signal or noise?
Does this story matter, or is it hype? Decide before you see what everyone else thinks.
Sources
- InfoQ