Real-Time Retail Intelligence: Building E-Commerce Recommendations with Lakebase and AI Search on Databricks

A leading fashion e-commerce platform in Asia implemented a real-time recommendation system on the Databricks Data Intelligence Platform to convert user intent signals—such as searches, product views, and cart interactions—into personalized product suggestions within milliseconds. The system processes approximately 1,000 events per second, including product views, searches, and purchases, using Lakeflow Connect’s Zerobus Ingest to land data directly into Unity Catalog Delta tables without requiring a self-managed message broker.

The architecture supports two serving paths: Path A pre-computes recommendations nightly for predictable surfaces like homepage carousels and category pages, while Path B handles dynamic, session-aware interactions such as 'Similar items' suggestions or real-time search re-ranking. Path A writes ranked product lists to Lakebase online tables, enabling the app to retrieve recommendations via a simple key-value lookup at serving time, avoiding model inference latency.

Path B processes real-time user signals sent directly to a Model Serving endpoint, executing a multi-stage pipeline—including vector search, LightGBM inference, and business rule application—within a single request-response cycle. The endpoint uses a custom MLflow PyFunc model to orchestrate the pipeline and applies commercial rules dynamically without redeploying the model.

The system includes fallback strategies for resilience, such as degrading to cached popular items if real-time latency exceeds thresholds, and handles new users and products by generating default embeddings from demographic signals or product attributes. Models are retrained weekly with MLflow tracking, and traffic is gradually shifted between model versions based on online performance metrics and business KPIs like conversion rate and revenue per session.