Databricks introduces FILE type for multimodal data

Databricks is releasing the beta of FILE type, a new column type designed to store unstructured data like documents, images, and videos natively within lakehouse tables. This allows unstructured data to be managed, secured, and queried alongside structured data.

FILE type enables the application of fine-grained access controls and security policies to unstructured files, mirroring those applied to standard tables. Deleting a row containing a FILE automatically deletes the corresponding binary in object storage, facilitating compliance with regulations like GDPR's 'right to be forgotten'.

Users can now run standard SQL and Python User Defined Functions (UDFs) directly on unstructured files stored as FILE type columns. This also supports building materialized views for incremental AI function execution.

FILE columns store lightweight pointers to file binaries rather than the binaries themselves, optimizing query performance. Actual file content is processed only when explicitly required by a query, preventing performance degradation from large data volumes.

Databricks is collaborating with the community to integrate FILE type support directly into Parquet and Delta Lake. This aims to ensure the entire ecosystem can leverage this capability, promoting open and portable multimodal data.

FILE type enables applications for enterprise document assistants, visual inspection of product images, call analytics from audio recordings, video understanding, and multimodal retrieval for AI agents. It streamlines the lifecycle from exploration to model training and real-time serving.