1. Yes, 1,9M rows is quite a lot for a dimension table (especially if you are going tenfold in the US). If you are going to use SCD type 2, this can get even worse.

2. Snowflaking will not help, as you already concluded.

3. This type of modelling is actually not that uncommon and is called mini-dimensions[/url]. They are well documented by Kimball and can reduce the complexity of your set-up enormously.

However, as you already mentioned - the link between the actual dimension and the mini dimension is through the fact table.

This can lead to other issues, such as dresses that have not sold anything. How can you analyze on dresses that have not sold a single piece, because there is not a transaction?

As mentioned in the article I linked to, you can create a factless fact table or something similar to track all the possible products.