Perplexity has pulled Amazon DynamoDB out of the read path that serves web content for its search product, replacing it with CobbleDB, an in-house key-value store of roughly 40,000 lines of Rust. Two engineers built it in about two months while directing hundreds of persistent, always-on coding agents, according to the company's published research on CobbleDB. None of those agents was permitted to authorize a production deployment.
Key takeaways
- Median batch-read latency fell from 31.4 ms on DynamoDB to 5.60 ms on CobbleDB, an improvement of roughly 82%, while 99th-percentile latency dropped from 123 ms to 24.2 ms.
- Two engineers directed hundreds of coding agents over about two months, retaining architecture decisions, code review and production authorization for themselves.
- Perplexity CEO Aravind Srinivas said migrating off DynamoDB could save the company up to $100 million annually, with internal projections putting storage-layer costs more than 20% below DynamoDB at scale.
Why Perplexity walked away from DynamoDB
The trigger was a combination of price and control. Perplexity concluded it was overpaying for DynamoDB while still not getting the tuning latitude it wanted over read performance, a constraint that matters when every answer the product generates depends on fetching many stored documents at once.
DynamoDB prices largely by bytes moved, a model that fits transactional workloads more comfortably than the wide, read-heavy fan-out patterns of an answer engine. Perplexity's argument is that a purpose-built store mapped to its own access shape could beat a general-purpose managed service on both axes at the same time.
How the system is put together
Rather than build one monolith, Perplexity split the job into three components that can be tuned independently. A component called Pillar manages document durability and tracks versions as records change. Lorry groups incoming updates into batches and moves them into the system. CobbleDB itself sits at the serving layer, dedicated entirely to low-latency reads at query time.
Underneath, CobbleDB leans on RocksDB for actual data transactions. On top of that foundation the team layered tunable partitioning and caching strategies aimed at holding performance during peak load. The published figures show the gains were consistent across the distribution, not just at the median: 90th-percentile latency fell from 56.7 ms to 9.77 ms.
What the agents were and were not allowed to do
The staffing story is the part the industry will argue about. Perplexity described the build as two engineers plus hundreds of proactive, always-on AI agents that supplied continuous inspection and follow-through across sessions. Srinivas framed the result on X as a DynamoDB replacement built by two engineers and hundreds of persistent agents over two months. The humans kept the decisions that carry blast radius.
That split is a deliberate governance line rather than a technical limitation. Agents generated and revised code; engineers owned the architecture, reviewed what landed, and held the authority to put anything in front of live traffic. Writing infrastructure and running infrastructure were treated as separate privileges.
How it compares to OpenAI's Rust rewrite
The shape will look familiar. Days earlier, OpenAI described two engineers and Codex rewriting the service behind every ChatGPT data read in Rust. Two labs independently landing on the same pattern β a very small human team, a large agent fleet, Rust, and a storage hot path β suggests the pattern is becoming a template rather than a one-off.
What Perplexity gave up
CobbleDB is not a DynamoDB feature clone. It does not support strong consistency or complex transactions, both of which a managed service provides by default. That is a workable trade for serving cached web documents to a search index, and a disqualifying one for systems that need transactional guarantees.
Perplexity has committed to open-sourcing CobbleDB once the system demonstrates production validation at scales of hundreds of thousands of requests per second, though it has not named a release date. Until code is public, the latency and cost figures rest on the company's own measurements.
FAQ
Is CobbleDB open source?
Not yet. Perplexity has said it intends to open-source CobbleDB after the system proves itself in production at scales of hundreds of thousands of requests per second. No release date has been announced.
Does CobbleDB replace DynamoDB entirely?
It replaces the read-serving layer for web content in Perplexity's search stack, not every DynamoDB use case. CobbleDB lacks strong consistency and complex transaction support, so workloads needing those guarantees are not candidates for it.
How much of the code did AI agents actually write?
Perplexity has not published a line-level breakdown of the roughly 40,000 lines of Rust. The company credits hundreds of persistent coding agents with the bulk of the implementation work across two months, with two engineers owning architecture, review and deployment authority.
