Except, oops, your user got married and moved, now your precious "referential integrity" means jack because the generated invoice is flat-out wrong. Product removed from the store? Too bad, needs to stay in the database forever for historical purposes. Prices need to change? Better design the database to handle snapshots of every product state.
If you were implementing this in MongoDB, you'd probably store a UUID and the flattened data at the time of invoice generation, that way you can still query on ids AND not deal with the headache of having a combinatorial explosion of data in your RDBMS.
You can then select fields from that JSON for invoices, reports, etc with the arrow operator:
https://www.postgresql.org/docs/9.6/static/functions-json.ht...
I feel a lot of people know that typical nosql databases (without database level transactions) are not suitable for their problem but they don't know why and then just think NoSQL is always bad and RDBMS are always better because the NoSQL databases are intended to be used for different problems.
But I do think having any data consisting of, say, items, orders, users, payment in MongoDB is very much a bad idea. Been there.
How would you represent a simple invoicing system in MongoDB (e.g. Customers + Products + Orders + OrderLineItems )? NoSQL-for-everything advocates posit two solutions: either denormalize the data by embedding Customer information within an Order document, which also contains an array of OrderLineItems, or use a UUID as a kind-of foreign key and maintain separate relationships. Both approaches have serious problems (data-duplication and inevitable inconsistency in the first, and lack of referential integrity in the second, besides ending-up abusing a NoSQL database as an RDBMS). Is there a better way? Or would you agree that certain classes of problems are best left to RDBMS' domain?