Perplexity develops CobbleDB, reducing batch-read latency by 82%
Summary
Perplexity has developed CobbleDB, a key-value database designed to enhance web content delivery for its search platform, with the core infrastructure built by two engineers and a team of proactive AI coding agents in just two months. This innovative architecture optimizes performance by separating durable document state, update delivery, and serving, allowing system replicas to process updates independently. As a result, the database achieved significant improvements in batch-read latency, reducing median latency from 31.4 ms to 5.60 ms, and the company plans to open-source CobbleDB to support other teams in the AI search domain.
Analysis
CobbleDB: CobbleDB is a key-value database developed internally by Perplexity specifically to serve web content for its AI search engine. Built with a focus on fast batch reads of prepared page records, it uses RocksDB MultiGet for parallel queries across partitions and same-zone replica routing for resilience. The system was created to replace DynamoDB and forms a core part of Perplexity's custom infrastructure stack alongside Pillar and Lorry. DynamoDB: DynamoDB is Amazon Web Services' fully managed NoSQL database service used for scalable key-value and document storage. Perplexity has migrated its search infrastructure away from DynamoDB to the new CobbleDB system to better align with workload patterns involving repeated batch reads. The replacement forms part of a broader architectural shift separating state, updates, and serving functions. YTsaurus: YTsaurus serves as the storage backend in Perplexity's Pillar component for maintaining versioned document state and queuing exports. It integrates into the company's data pipeline to feed partitioned updates to Lorry for S3 batch preparation before delivery to CobbleDB replicas. This setup enables independent and ordered ingestion across replicas without blocking recovery or scaling. Perplexity: Perplexity operates an AI-powered search platform that delivers conversational answers with cited sources. The company built CobbleDB to optimize performance and costs for its high-volume search workload, separating concerns across durable document storage, update delivery, and serving layers. Perplexity plans to open-source the database to support other teams developing AI search systems. Architecture: Perplexity separated durable document state in Pillar, update delivery via Lorry, and serving in CobbleDB to allow replicas to ingest updates independently at their own pace. Open Source Plans: Perplexity intends to release CobbleDB as open source to benefit other teams building AI search applications. Infrastructure Development: Perplexity built CobbleDB using two engineers directing a team of proactive AI coding agents for continuous inspection, follow-through, and architecture validation over two months.
Categories
techai_agentsaimachine_learning