Introducing FILE type: a native column type for multimodal data

Why it matters

Databricks has launched a beta of FILE, a column type that stores unstructured content such as documents, images, and video directly in lakehouse tables. Authors Michael Armbrust, Burak Yavuz, Dejan Krakovic, and John Spencer frame the feature as a way to make unstructured data queryable, secured, and managed alongside structured data rather than in a separate system. The column holds a lightweight reference, so the engine retrieves actual bytes only when needed, avoiding the performance cost of embedding raw binary in tables. FILE integrates with Unity Catalog, applying row- and column-level access controls plus attribute-based access control, which addresses the governance gap teams face when storing file paths as plain URL strings governed by folder-level permissions. Lifecycle management is also tied to rows, so deleting a row keeps data and references in sync, helping with compliance needs like GDPR. Databricks is developing the feature in the open through Parquet and Delta Lake, aiming to keep multimodal data portable across vendors and model providers.