Perplexity unveiled CobbleDB on September 14, a custom key-value store written in about 40,000 lines of Rust. It replaces the company’s previous Amazon DynamoDB setup for the hot data that powers its AI search. Two engineers directed hundreds of internal AI coding agents and finished the core system in roughly two months. Median batch-read latency fell from 31.4 ms to 5.60 ms, an 82 percent drop. The 90th percentile moved from 56.7 ms to 9.77 ms and the 99th percentile from 123 ms to 24.2 ms. The design splits work across three pieces: Pillar for durable document state, Lorry for batched updates, and CobbleDB itself for fast serving. It sits on RocksDB with tuned partitioning and caching aimed at heavy parallel reads. The store skips strong consistency and complex transactions. Perplexity projects more than 20 percent lower storage costs at scale and plans to open-source the system once it handles hundreds of thousands of requests per second in production.
Perplexity unveiled CobbleDB on September 14, a custom key-value store written in about 40,000 lines of Rust. It replaces the company’s previous Amazon DynamoDB setup for the hot data that powers its AI search. Two engineers directed hundreds of internal AI coding agents and finished the core system in roughly two months. Median batch-read latency fell from 31.4 ms to 5.60 ms, an 82 percent drop. The 90th percentile moved from 56.7 ms to 9.77 ms and the 99th percentile from 123 ms to 24.2 ms. The design splits work across three pieces: Pillar for durable document state, Lorry for batched updates, and CobbleDB itself for fast serving. It sits on RocksDB with tuned partitioning and caching aimed at heavy parallel reads. The store skips strong consistency and complex transactions. Perplexity projects more than 20 percent lower storage costs at scale and plans to open-source the system once it handles hundreds of thousands of requests per second in production.