Perplexity swapped out DynamoDB with two humans and hundreds of agents

Median batch-read latency went from 31.4ms to 5.60ms. p99 went from 123ms to 24.2ms. The internal cost model says at least 20% savings against DynamoDB, and Aravind Srinivas puts the annual figure at up to $100 million. Those are the numbers in Perplexity's CobbleDB write-up, and they're the least surprising part.

CobbleDB is a key-value hot store for the web content backing Perplexity's search index. Architecturally it's not exotic: RocksDB underneath, MultiGet with keys grouped by partition and read in parallel. Anyone who's tuned RocksDB will recognize the shape. The part I had to read twice is the staffing. Two engineers. Two months. About 40,000 lines of Rust. Hundreds of persistent agents doing the inspection and follow-through work across sessions, with the humans holding architecture, code review, and production authorization.

I've built this kind of thing. Not at Perplexity's scale, but I've written the storage layer that sits between a search index and the thing serving it, and the work is 20% design and 80% grinding through edge cases, benchmarking, fixing the benchmark, re-benchmarking. The grind is exactly what a persistent agent fleet is good at, and it's exactly what burns senior engineers out. The division of labor here is the copyable artifact: humans own the three decisions that are expensive to reverse (what the system is, whether the code is correct, whether it goes to production), agents own everything between those decisions.

What I don't know from the post is the failure rate. Hundreds of agents over two months generates a lot of discarded work, and Perplexity doesn't publish how much. A 40,000-line result could sit on top of 400,000 lines of rejected attempts, and that's a different economics story than the headline. The latency table is real and measured. The staffing model is described, not measured.

The thing to take from this if you run infra: price your managed-service line as a build decision again. For years the answer was obviously "don't build your own database," because the engineering cost dominated the bill. Perplexity's claim is that the engineering cost moved. You don't need to believe the $100M figure to run the arithmetic on your own DynamoDB or DocumentDB spend against two engineers and two months. That calculation has a different answer than it did in 2024, and the write-up gives you enough detail (partition-grouped MultiGet, hot-store-only scope, humans on the review gate) to sanity-check whether your workload fits the shape.

Note what CobbleDB isn't. It's a hot store for one access pattern, not a general-purpose database. The scoping is what makes the two-month number credible.