I think the problem is fundamentally that if you take a record, and split it across 2 tables, then there's no easy operation to merge them and guarantee distinctness.
Separately, even though there are some %-breakdowns in the article, I don't think the `DISTINCT` sort is so much of a problem in most cases. Usually the actually joining is the problem for any truly large dataset, so while I trust the diagrams shown, I think those %'s change dramatically based on the underlying data.