Details
- Perplexity says it is publishing research on CobbleDB, the key-value database it built to serve web content for Perplexity Search.
- The system was built in about two months by two engineers plus a large swarm of proactive AI coding agents.
- CobbleDB is designed for repeated batch reads of prepared page records, not as a general-purpose database.
- Its architecture groups page keys by partition, reads them in parallel through RocksDB MultiGet, and routes around slow replicas.
- Perplexity split the storage stack into Pillar for durable document state, Lorry for batched update delivery, and CobbleDB for serving.
- The company says this separation lets each replica ingest updates independently, so recovery on one node does not slow others.
- After replacing DynamoDB, Perplexity reports median batch-read latency dropped from 31.4 ms to 5.60 ms and p99 from 123 ms to 24.2 ms.
- Perplexity estimates the new stack delivers at least 20% savings versus DynamoDB and plans to open-source CobbleDB for other AI search teams.
Impact
CobbleDB shows how AI-search companies are increasingly treating storage as a product advantage, not just an infrastructure cost center. The reported latency and cost gains suggest Perplexity is tightening the economics of high-volume retrieval while using AI agents to compress development cycles. If it open-sources the system, it could give other AI search builders a practical reference for read-heavy serving architectures, potentially pressuring teams still relying on more general-purpose databases.
