When analyzing a large build found the size of the refs table and its indexes dwarfed the rest of the database:

Looking at the table its clear that there is a huge amount repeated string content in this table.
The most important data is object, referenced_object columns. But tracking the property_path and property_type as strings would greatly bulk up the size.

Deduplicating those strings into side tables would complicate some queries but probably make a huge different in database size and potentially improve analyze times.