O’Reilly Media

Multimodal Lakehouse: Evolving Data & Workloads with Chang She

In the talk, Chang She outlines a “multimodal lakehouse” that extends traditional data‑lake architectures to handle not only tabular records but also embeddings, images, video and other rich media. He explains three shifts: first, data formats now span multimodal types, demanding new read/write engines and metadata synchronization. Second, workloads move beyond pure SQL analytics to include vector search, model training, and GPU‑intensive feature engineering. Third, conventional lakehouses excel at batch processing but require separate OLTP systems for real‑time serving; the multimodal lakehouse aims to unify both. She cites the Last DB format as a concrete innovation that lets the same storage layer support online and offline queries simultaneously, eliminating the need for a separate serving database. This unified approach simplifies pipelines that ingest embeddings, run similarity searches, and feed results into downstream AI models. For enterprises, the ability to store, process, and serve multimodal data from a single platform can cut infrastructure costs, accelerate AI product cycles, and open new revenue streams that rely on real‑time, media‑rich insights.