If you can denormalize your schema so there's no joining, that's great. If you can change the product so there's no joining, great. If the schema and database are designed to do the join you're going to do, that's probably fine.
If not, it's cheaper and easier to add capacity to the pool of shared nothing web machines than to add capacity to the pool of databases machines, so grab the data and run the join on the webserver. (or at least benchmark it; you may be surprised)
The lesson is that if you find yourself doing joins, then you need to consider whether the database should be redisgned. A join is sometimes necessary, and the engine is the right place to do it if it must be done, but it's a warning. So if possible, design your DB to avoid joins.
Disagree. If you need to join data, you have do do it somewhere. You think you are going to do a better job of it than the database engine, which has probably been optimized for years if not decades to do this very thing?