Perplexity Dramatically Reduces Query Latency with In-House Developed Database

Perplexity, an AI search service provider, has completed its migration from Amazon's DynamoDB database service to CobbleDB, a proprietary key-value data store developed in-house. CobbleDB is developed in Rust, a programming language that significantly reduces query latency when processing large volumes of documents in batch while also achieving cost reductions in cloud storage.

As a rapidly growing AI search service, Perplexity processes massive search queries daily. As user numbers increase, database response speed and cost efficiency have become critical factors directly affecting service quality, and pre-built cloud services alone have proven insufficient for certain scenarios. Against this backdrop, the company decided to develop an internally-optimized database tailored to its own workloads.

CobbleDB is a key-value data store specialized in rapidly reading and writing values corresponding to specific keys. Rust was chosen for its development, offering both memory management safety and processing speed—key characteristics of the language. Through this migration, query latency has improved in operations handling large documents, and storage cost reductions have been confirmed. The company explains that this new architecture can efficiently handle large-scale production traffic.

Infrastructure internalization is an approach pursued by major tech companies, but few companies at the startup stage have undertaken it comprehensively. Perplexity's decision to switch its database layer to in-house development at this growth stage demonstrates that the company prioritizes balancing rapid scaling with service quality. While pre-built cloud services excel in ease of deployment, custom implementations can offer performance and cost advantages in specific processing patterns—a principle exemplified by this migration.

In the AI search market, response speed is a critical metric directly impacting user experience. Since AI model inference processing is added to search result generation, database layer latency typically becomes a bottleneck. Reducing query latency can positively affect the search speed ultimately experienced by users.

Key points for future observation include whether CobbleDB can operate stably at even larger traffic scales. While in-house databases offer high customization flexibility, the company must continuously build internal maintenance and operational expertise. Whether Perplexity continues down this path of infrastructure self-development or adopts a hybrid approach using existing services for other domains will be an important indicator for understanding the company's engineering strategy.