Inside the feature store powering real-time AI in Dropbox Dash

Dropbox Dash built a custom hybrid feature store to power real-time AI ranking across tens of thousands of work documents.

• •Infrastructure spans two environments (on-premises low-latency and cloud-based Spark), ruling out standard cloud-native feature stores
• •Feast was chosen as the orchestration layer; its Python serving path was replaced with a Go service achieving p95 latencies of 25-35ms
• •Dynovault (in-house DynamoDB-compatible storage) co-located with inference workloads delivers ~20ms client-side latency
• •A three-part ingestion system (batch, streaming, direct writes) keeps features fresh, with change detection cutting batch update times from 1+ hour to under 5 minutes
• •Only 1-5% of feature values change per 15-minute window; targeting that subset reduced write volumes from hundreds of millions to under one million records per run