Source: Pinecone serverless architecture (Pinecone blog)

Pinecone's description of its serverless design: blob storage as source of truth, geometric partitioning, a freshness layer, namespaces and cached compute.

Contents

Summary

Pinecone serverless separates storage from compute. Writes go to a log and to blob storage, which is the source of truth. An index builder partitions vectors geometrically around centroids into immutable segments, while a small freshness index covers recent writes so they are searchable within seconds. Query executors search both and cache segments locally.

Key claims

  • Blob storage holds every index; compute loads only the segments it needs, so the two scale independently.
  • The index builder keeps a coarse-to-fine hierarchy of geometric partitions and rebuilds incrementally.
  • A freshness layer makes inserts, updates and deletes searchable within seconds (a lambda-style architecture).
  • Namespaces are hard partitions and the unit of tenant isolation; an index can hold hundreds of thousands of them, with cold namespaces paged in on demand.
  • Cold-start (first query) latency is a couple of seconds for most datasets and up to about 20 s at billion scale.

Feeds

Vector databases

Suggest changes →