Feature Store for Real-time AI

1 post · Oct 2025 – Dec 2025

What this trail covers

Dropbox developed a hybrid feature store for its Dash AI product to power real-time ranking and AI-driven search. The system integrates on-premises and cloud environments, leveraging Feast for orchestration, Spark for computation, and a custom Go serving layer with Dynovault (DynamoDB-compatible) for low-latency online lookups. It employs a multi-part ingestion strategy (batch, streaming, direct writes) to balance feature freshness with computational efficiency and infrastructure load, enabling real-time AI capabilities. The Go serving layer replaced a Python implementation to achieve sub-100ms latency targets, and intelligent change detection in batch ingestion reduced update times from over an hour to under five minutes.

All Dropbox trails

Dec 2025Inside the feature store powering real-time AI in Dropbox DashThis post details the evolution of Dropbox's feature store for Dash AI, focusing on architectural choices and performance optimizations. Key contributions include: replacing a Python feature serving layer with a Go service to achieve sub-100ms latency targets, leveraging goroutines and optimized JSON parsing for true concurrency; implementing a three-part ingestion system (batch, streaming, direct writes) to balance feature freshness with computational efficiency, including intelligent change detection in batch ingestion to reduce update times from over an hour to under five minutes; and integrating Feast for orchestration, Spark for computation, and Dynovault (DynamoDB-compatible) for low-latency online serving. The post highlights tradeoffs made in choosing Feast and designing the hybrid architecture to bridge on-premises and cloud environments.
Oct 2025With Mobius Labs' Aana models, we're bringing deeper multimodal understanding to Dropbox DashThis post details the integration of Mobius Labs' Aana multimodal models into Dropbox Dash. Aana provides an ultra-efficient architecture for understanding rich media (audio, video, images) at scale, analyzing how different modalities relate to each other. Key technical components include inference-optimized Whisper-like models, transformer-based and MoE vision/language models, and advanced inference optimizations like HQQ (Half-Quadratic Quantization) for low-bit inference and Gemlite for accelerating AI operations. The Aana SDK orchestrates these models for batching, coordination, and GPU utilization, enabling the analysis of exabytes of content with reduced computational requirements. This enhances Dash's ability to search and interpret complex scenes, recognize objects/actions, and provide context-aware insights.
Advertisement