Perplexity’s search stack now runs on CobbleDB, a Rust‑written key‑value store that the team built to replace Amazon DynamoDB. The swap delivered up to a five‑fold drop in query latency and lowered the amount of cloud storage needed for large document batches, directly affecting throughput and cost for any AI‑driven search workload.
Why the change matters to engineers
AI and ML pipelines often need to fetch millions of vectors or metadata records quickly. A reduction in latency translates to faster response times for end‑users and tighter SLAs for downstream services. Lower storage consumption reduces the bill for hot data that must be kept readily accessible, a common pain point for teams scaling search indexes.
Architectural and implementation considerations
CobbleDB is a self‑hosted, Rust‑based key‑value store, meaning the responsibility for provisioning, scaling, and maintaining the service shifts from a managed DynamoDB offering to the Perplexity operations team. The move implies:
- Control over data layout and on‑disk format, which can be tuned for the specific access patterns of search queries.
- Need to implement replication, backup, and disaster‑recovery mechanisms that DynamoDB provides out‑of‑the‑box.
- Potential to integrate Rust‑level performance optimisations directly into the storage engine.
Operational and security implications
Running a custom store introduces new operational load: monitoring latency, storage health, and capacity planning become internal tasks. Security teams must ensure that data at rest is encrypted, that network access is restricted, and that any authentication layer is correctly configured, because the built‑in IAM integration of DynamoDB is no longer present.
Related CloudNinjas coverage: hands-on guides.
What This Means For Practitioners
Teams should evaluate whether the performance and cost gains of a bespoke store outweigh the added operational overhead. Start by instrumenting query latency and storage metrics on the new stack, verify backup and recovery processes, and review access controls to match or exceed the security posture previously provided by DynamoDB.