Why Perplexity Built Its Own Database to Beat AWS

Perplexity replaced Amazon's DynamoDB with a custom database, cutting query latency by five times and reducing costs.

AI search company Perplexity has made a significant shift in its core infrastructure, replacing Amazon's widely used DynamoDB with a custom-built key-value store named CobbleDB. According to a report from InfoQ, this strategic move was driven by the need to handle large batches of documents more efficiently for its search product. The new system, developed in-house using the Rust programming language, has delivered impressive results. Perplexity successfully cut its query latency by a factor of five and also achieved a notable reduction in cloud storage costs, all while managing the heavy demands of its production traffic.

While specific architectural details of CobbleDB remain internal, the decision to build a custom solution points to limitations in general-purpose managed databases when faced with highly specialized workloads. DynamoDB is a powerful and scalable service, but its performance and pricing models are designed for a broad range of use cases. Perplexity's search infrastructure likely requires extremely fast, low-latency access to massive datasets, a scenario where the overhead of a generic service can become a bottleneck. By building CobbleDB in Rust, Perplexity's engineers could tailor every aspect of the database to their exact needs—from data serialization and indexing strategies to memory management. Rust's performance characteristics, particularly its lack of a garbage collector, would prevent unpredictable latency spikes that are unacceptable in a real-time search application, ensuring a consistently fast user experience.

Perplexity's move is a modern chapter in the long-running "build versus buy" debate for critical infrastructure. While the cloud era has made it easier than ever to "buy" infrastructure components as managed services, high-growth companies often reach a tipping point where the trade-offs no longer make sense. At a certain scale, the performance constraints and recurring costs of off-the-shelf services can stifle innovation and hurt margins. This pattern has been seen before, with tech giants like Meta and Dropbox eventually building their own custom storage and database systems to escape the limitations of third-party vendors. What makes Perplexity's case notable is that it's happening in the AI-native generation of companies, suggesting that the intense performance and cost-efficiency demands of AI models may accelerate this trend of moving from generic cloud services to bespoke, highly optimized infrastructure.

The case study serves as a crucial reminder for CTOs, founders, and engineering leaders to continuously re-evaluate their infrastructure stack. Managed services are invaluable for speed and initial scale, but they are not a one-size-fits-all solution for the entire life of a company. It's essential to understand the specific performance bottlenecks and cost drivers of your core application. The decision to build a custom system like CobbleDB is a major undertaking that requires deep systems engineering expertise and a commitment to long-term maintenance. However, as Perplexity has shown, for companies where latency and operational efficiency are direct competitive advantages, the strategic investment in building a tailored solution can unlock performance and cost savings that are simply unattainable with generic tools. The key is knowing when your scale and specific needs justify taking that leap.

This story challenges the default 'use a managed service' wisdom. It shows that for performance-critical AI applications, building a custom data store can yield order-of-magnitude improvements in latency and cost, providing a significant competitive edge over those relying on generic cloud infrastructure.

Perplexity's move directly translates to a better user experience and a lower operational cost base. By reducing latency and cloud bills, the company improves its product's competitiveness and gross margins, allowing for reinvestment in R&D. This is a strategic infrastructure decision with direct financial benefits.