In the mongodb source, when NOT using the randomSampling, would it make sense to actually pull out the N latest documents, and not the oldest? (or at least have an option to reverse the order) My thinking is that in long-lived collections the later documents might be more representative of the "schema" at this moment. And there might also be a tiny performance advantage for huge collections (in-memory vs on disk)?

Yup that makes sense.

Looking at the code, I realize that the document size labeling might be a bit of a resource hog with our +700 Gb collections.