Perplexity builds CobbleDB in two months, cutting batch-read latency by 82%

Perplexity unveiled CobbleDB on September 14, 2026, a custom key-value hot store written in approximately 40,000 lines of Rust. The system was purpose-built to handle the specific read-heavy demands of AI-powered search, and it replaced a prior architecture built on Amazon DynamoDB. What the numbers actually say Median batch-read latency dropped from 31.4 ms to 5.60 ms, an improvement of roughly 82%. The 90th-percentile latency fell from 56.7 ms to 9.77 ms. At the 99th percentile, latency dropped from 123 ms to 24.2 ms. Cost is the other side of the ledger. Perplexity’s internal projections suggest CobbleDB could reduce storage-layer costs by more than 20% compared to DynamoDB at scale. The savings come primarily from eliminating DynamoDB’s per-byte pricing model and replacing it with workload-specific optimizations that map more cleanly to how AI search actually behaves. How the architecture works CobbleDB follows a decoupled design with three distinct components. Pillar manages document durability and handles updates. Lorry handles batched ingestion, acting as the pipeline that moves data into the system efficiently. CobbleDB itself sits at the serving layer, dedicated entirely t...