Perplexity’s AI agents helped build a database. They weren’t allowed to run it.
The New Stack Amanda Caswell
Perplexity built its own database with AI agents, but kept the bots off the controls. It’s faster and cheaper than DynamoDB on paper, which is the part that stings for cloud services.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Perplexity got fed up with DynamoDB’s cost and its lack of control over read performance, so it built CobbleDB instead. Two engineers did the work in two months, with hundreds of coding agents helping along the way. The result is a roughly 40,000-line Rust key-value store that now serves part of Perplexity’s production search traffic.
The company says the new setup cut median batch-read latency to 5.6 milliseconds from 31.4 milliseconds on DynamoDB, with p99 dropping from 123 milliseconds to 24.2 milliseconds. Those numbers were recorded before and after the cutover, not in a side-by-side test, so they’re useful but not perfect. Still, the direction is clear enough: Perplexity’s search stack got much faster, and the company expects CobbleDB to cost at least 20% less than DynamoDB. It also plans to open-source the database later.
The reason DynamoDB started to look awkward is buried in how Perplexity search works. Each Search API call can fetch 100 to 120 page keys in batches of 10 to 20, and each item averages about 50 KB. A slow replica could hold up the whole batch, and DynamoDB gave Perplexity little room to tune read behavior. On top of that, the steady stream of large reads and writes from search, crawling and reprocessing made the cloud bill harder to defend as traffic and the corpus grew.
Perplexity split the storage system into three layers to deal with that. Pillar keeps durable document state in YTsaurus on HDDs, Lorry packages updates and moves them through S3, and CobbleDB handles processed page data across three replicas per partition. Reads stay in the same availability zone when possible, the router can skip a slow replica, and updates can land independently so one replica can lag without freezing the others.
The synthetic and load tests are eye-catching too. Perplexity says CobbleDB reached 500,000 requests per second before performance started to decline, and its own benchmark used batches of 10 to 15 keys with values from 100 bytes to 100 KiB. The bigger story, though, is that AI agents helped build the system, but the two engineers kept the keys. That feels like the right division of labour when the data matters and the rollback button is more fantasy than feature.
My take — AI-written commentary, not fact-checked reporting
This is the sane version of AI-assisted infrastructure: let the agents write, test and grind through the boring parts, then keep them well away from production. Database work is where hype goes to get humbled, and Perplexity clearly understood that part. The real takeaway isn’t that agents can run a database; it’s that they can make bespoke systems cheaper to build, which is much more useful and much less stupid.
Read more about this at: The New Stack