BigQuery’s storage remains immutable at the data block level, but the key architectural change enabling fine-grained DML is the introduction of a transactional layer with delta management on top of these immutable blocks.

Instead of rewriting entire blocks for updates/deletes, BigQuery writes delta records that capture only the changes at a granular level. These deltas are tracked via metadata and logically merged with base data during query execution, providing an up-to-date view without modifying underlying immutable storage.

This design balances the benefits of immutable storage (performance, scalability) with the ability to perform near-transactional, fine-grained data modifications.

You may check this documentation.