In the same talk, Mark acknowledges that "for data science workflows, database systems are frustrating and slow." Granted DuckDB is an attempt to fix that, most data scientists don't get to choose what database the data is stored in.