1. Denormalized data is often a boon. One of your key use cases is "give me all the information about this Product so I can show it to the user." Forget joining a ton of tables, just do 1 read of 1 product and go on your way.
2. A certain amount of more freeform data is expected. Different product categories will have different sets of information. T-shirts and drinkware both have sizes, but these sizes have nothing to do with each other. You can model all this in a traditional SQL database, but you have to stop and think really hard about it, and potentially end up with a plethora of tables. In a document database, it's much easier to just add the data and go on your way.
I think being able to convert back and forth would make it so worthwhile!
What's the alternative you want to move to?
The biggest issue right now is that MongoDB compares and sorts values differently than PostgreSQL/jsonb. For that reason, we have to do a lot of filtering on the FerretDB side, and that can't be great for performance. Pushing more work on the database size should make FerretDB perform much better.
I wouldn't go into the document vs. relational argument, all arguments for and against would have merit. There are valid use cases for document databases (take e-commerce, for example), and we should not discount the fact that using a relational database is just more complicated. Using vanilla Postgres for a MongoDB use case will not be feasible for someone who's focus is, let's say, mobile application development. There is a reason behind MongoDB's popularity - it just provides a great developer experience. This is what we are aiming to recreate on top of Postgres.
[1]: https://blog.ferretdb.io/pjson-how-to-store-bson-in-jsonb/