And if API matches data format closely, we won't really need to allocate copies of data elements on heap,
We only keep a copy of the number of rows per table and some schema-related data in metadata_table_info for convenience, but I'll see if I can also reduce those allocations.