Let us for the explanation say we have three tables: types, articles, article_translations. The first table is clear what it includes. It includes an id, and description. The article will contain all non translatable data of an article and article_translations all translations of an article.
So now you place an repeater on a page. The dataset would be article_translations, for you want to display translatable text and non translatable text. The filter would be the language. Now on the elements of the item you can chose which columns to use. You will see information from the article and article translation table. However, you cant chose data from the types table. Since the article is referenced to type, by id, you can't display the text value of the type in the itme.
Make it possible to chose from all referenced tables. This does mean more load. So I can expect this will not be done. But for then, for the more advanced people, the next option would be nice.
You can set repeater data with the result of a query. However, when you for example make a query on 'article_translations' and you log the result.items you will will see all columns of that table and a reference column. It would be nice that it will return all referenced columns. Technically it is possible to exclude columns you don't want to pass to the client, so this way you can still reduce the amount of data. This will prevent having to make multiple datasets.
Offcourse, people going to use this will also need to keep performance in mind. If you don't filter your data correctly, or just return to a client what you need, there is a lot of data which may go to the client.