Perplexity Cuts Search Latency by 82% With CobbleDB

Perplexity has migrated its primary web-content retrieval path from Amazon DynamoDB to a custom-built key-value store called CobbleDB. Designed to handle large batches of prepared page records, the new database has dramatically improved performance under live production traffic of approximately 200,000 requests per second. According to Perplexity, median batch-read latency plummeted from 31.4 milliseconds to just 5.60 milliseconds, representing an 82% reduction. Additionally, tail latency at the 99th percentile dropped from 123 milliseconds to 24.2 milliseconds, while internal cost projections indicate savings of at least 20% across all commitment tiers.

The system is optimized for Perplexity's specific retrieval patterns, where a typical search API request spans 100 to 120 page keys, grouped into batches of 10 to 20 keys with an average record size of 50 KB. To achieve these speeds, the architecture separates durable document state, batch delivery, and low-latency serving. Durable state is managed by Pillar on YTsaurus, while Lorry handles batch delivery to Amazon S3. CobbleDB itself runs on RocksDB, utilizing its MultiGet operation for batched fetches. A stateless router directs traffic to replicas in the same availability zone and hedges slow requests to prevent stragglers from delaying entire batches.

Remarkably, the project was completed in just two months by a team of only two human engineers working alongside hundreds of persistent AI coding agents. Together, they generated roughly 40,000 lines of Rust code and managed the surrounding migration. The AI agents assisted by auditing channels, reviewing code changes, and preparing tests. While CobbleDB requires developers to manage partitioning, replica health, and RocksDB compaction manually, Perplexity plans to release the database as an open-source project on GitHub, allowing the broader community to evaluate and adopt the technology.