Wow, @JaneW, my first new question of the week, and you asked a great one.
Xcalar's visual modeling tools were architected to work with structured, semi-structured and unstructured data. Because Xcalar works with metadata about your data, rather than directly with your data, XD brings all columns from your dataset through the modeling process with no significant performance penalties. the select columns later option is liberating. You can always add a column from your dataset into your table, to refine your insights whenever the need arises. Although you may have architected the entire data model before starting your modeling work with XD, it's not necessary.
The Select Columns Later option lets you:
* Start modeling organically, where step one is adding a column from your dataset, steps 2 - 10 might be manipulations of the column, and step 11 might be adding another column from the dataset.
* Cleanly refine semi-structured or unstructured data into structured data to seamlessly work with structured data.
Try it, you might like it!