We solve this by managing referential integrity across relational databases when we subset. Effectively you have to build a graph of edges and nodes where edges are constraints and nodes are tables, then you have to follow these edges from node to node to maintain the constraints. It's certainly not trivial and gets very complicated when people use sensitive data as a primary key (for ex. email)/foreign key, then you need a way to anonymize those emails and shelf those primary keys until you can re-insert them later so you don't break the constraints.

(disclaimer - co-founder of Neosync: github.com/nucleuscloud/neosync, open source tool to generate synthetic data and orchestrate and anonymize data across database)