Databricks Unveils Real-Time E-Commerce Recommendations with Lakebase and AI Search

Databricks has launched a unified platform for e-commerce recommendations, leveraging Lakebase and AI Search to enhance real-time personalization. The system ingests clickstream data at a rate of 1,000 events per second, enabling a potential 10–30% increase in conversion rates through effective personalization. It supports over 100,000 SKUs and typically returns 50–100 ranked products per user query with a latency of just 2-digit milliseconds for pre-computed batch recommendations.

The integration utilizes Zerobus as the ingestion backbone for clickstream data and employs LightGBM as the scoring model for predictions. Daily updates ensure fresh recommendations based on the latest inventory and user preferences, while behavioral aggregates are updated daily, and a full product catalog sync occurs weekly. Model retraining is also conducted weekly using Databricks Workflows, with new products surfacing in recommendations within 24 hours after initial embeddings.

Key performance metrics such as AUC-ROC, NDCG@K, Recall@K, and Log Loss are monitored to assess model performance. This initiative follows Databricks' recent enhancements in marketing ROI tools and data governance features, further solidifying its position in the e-commerce analytics space.