The Costly Illusion of Denormalization for Performance

My October post @All Analytics.

Be that as it may, practitioners insist that performance improves when they denormalize databases, because "bundling" facts into less relations reduces joins. But even if this were always true -- it is not -- performance gains, if any, do not come from denormalization per se, but from trading off integrity for performance. What many data professionals miss is that the redundancy introduced by denormalization must be controlled by the DBMS to ensure data integrity, which requires special integrity constraints that, it turns out, involve the very joins that denormalization is intended to avoid, defeating its purpose. These constraints are practically never declared and enforced, which creates the illusion that denormalization improves performance at no cost.

Read it all. (Please comment there, not here)