Perplexity has moved its main search-serving tier from Amazon DynamoDB to an internally built distributed key-value store called CobbleDB. Production measurements cited by InfoQ show median batch-read latency falling from 31.4 milliseconds to 5.6 milliseconds, while overall storage expense declined by at least 20 percent.
The company's answer engine creates an unusual database workload. A query can request 100 to 120 pages, with records averaging about 50 kilobytes because they include full text chunks and vector embeddings. At more than 200,000 requests per second, byte-based cloud pricing became expensive, and engineers had limited control over cache placement, replica selection and cross-zone reads.
The replacement separates durable data from fast retrieval. Pillar stores versioned source data, Lorry packages updates into files on Amazon S3, and CobbleDB workers load those batches onto local NVMe drives using RocksDB. A stateless router keeps reads in the same availability zone when possible and sends a backup request to another replica when one responds slowly.
CobbleDB deliberately accepts eventual consistency and asynchronous replication because search results can tolerate short update delays. That choice removes some coordination overhead, but Perplexity now owns node maintenance, backup validation and partition rebalancing that DynamoDB previously handled. The company says two systems engineers, assisted by coding agents, built roughly 40,000 lines of Rust in two months and plans to release the code as open source.