Never read this before. It seems absurd for a few reasons.
1) What if you use metadata to drive out measures? We have hundreds of measures in our document and we use a unrelated table to store the names of these measures, thresholds (for turning them green, yellow and red) and formulas for their calciulation. We also use meta data for security and these two sets of metadata are "isolated" islands i guess.
2) Another huge problem with this limitation is many to many relationships. By proper design you need to have a junction or bridge table between such dimensions. THis will break their cardinatlity descreases as you move away from fact table rule.
3) Finally, QV does not work well with proper "enterprise" DW with conformed dimensions. Take a simple example where you have two fact tables each joined to a conformed time dimension -> FactA, FactB and Time. Lets also say that only Facta is linked to a product dimension. If you choose a product it will filter out data from FactA. It will also filter out Time dimension members that have no data associated with the product you chose. Now here's the key. This will also filter out Time members associated with "real" fact that you will have in FactB. This is really a tough one for me. QV uses an inner join type of filter for this and unfortunately the really cool 'associative' model breaks down with enterprise conformed dimensional model. The solution is data island. Isolate your fact tables. THis is going back to a data mart silo approach but at least you data will be correct.
