How does Pinecone serverless separate storage and compute?

Pinecone serverless stores vector data in object storage like S3, while compute nodes load data on demand to serve queries. When a query arrives for a namespace or shard not currently in memory, the compute layer fetches it from storage, pays a cold-start latency penalty, and caches it for subsequent queries. This architecture scales storage near-infinitely at object storage cost, avoids idle compute fees, and handles spiky workloads without overprovisioning. It trades some tail latency for dramatically lower total cost at rest.