Why Multimodal
Data Needs a Better
Lakehouse?

We explore the challenges and limitations of current data lakehouse architectures in handling multimodal data, crucial to modern machine learning and AI workloads:

• Current lakehouses lack native support for unstructured data like images, audio, and video.
• AI and ML workloads depend on smooth handling of diverse, multimodal data types.
• A better lakehouse should unify storage, metadata, and fast access across all modalities.

We propose design principles and potential system enhancements for a new generation of multimodal lakehouses, aiming to bridge the gap between traditional data infrastructure and the needs of large-scale, AI-driven applications.