Yes, you can blend two data sources. My general rule is: when in doubt, try it!
While the source may contain many joins in a star schema, the resulting data set as seen in Tableau is a flat table. So, you would be simply blending two flat tables, regardless of the underlying structure that produced the tables.
I don't understand the "And why" at the end of question 1. Why is it okay to do this? Because you can. Because it accomplishes your goal.
As for the relationships box, you should remove any that are inappropriate (i.e. the names may be the same but they mean different things), and add any that are appropriate (i.e. equivalent fields that have different names).
As for understanding what is happening "behind the scenes," I think the best authority on data blending that I know is Tableau Zen Master Jonathan Drummey. I highly recommend reading articles in his blog site Drawing with Numbers | Thoughts on data visualization and Tableau, and watching the video from his TC14 session Post-Conference Materials | Tableau Conference 2014. In short, blending is somewhat similar to a SQL left-join, but it is different in that the right side of the "join" (secondary table in the blend) must be aggregated. There are a few other idiosyncrasies that Jonathan details.
I hope that helps.
I am in process of clarifying this with Tableau support. When I get the full understanding I will post the answer here.
Ok, after some discussions with Tableau support here is the answers:
Is it possible/okay (supported by Tableau)? - Yes it is possible/supported to do data blending on two tableau data sources which are star schemas of their own.
Why: mainly, the way I understood it, that every data source modelled in tableau, after all the joins are done is essentially a large flat table (conceptually), so blending two star schemas is like blending two huge flat data tables, where columns are all the columns in all the dimensions.
What to be aware/wary of when doing it:
1) left join: remember that blending is always a left join on the primary data source. So the data in secondary, that doesn't match, will be excluded
2) non-additive measures (such as count distinct, medians, etc) need special care
3) cardinality of the attributes on which the acting relationships are set. As for any blending, the smaller cardinality, the better performance. If you set blending relationships on high cardinality columns, good luck waiting for the results. And generally bad for performance.