“Q17 is the trophy: a correlated subquery averaging
l_quantityper part that, back when it evaluated once per outer row against 6M line items at scale factor 1, took 32.7 seconds. Fully pushed down, it’s 37 milliseconds.”
32.7 seconds to 37 milliseconds is what happens when you stop shipping 6 million rows across a wire to evaluate them one at a time. The scoreboard moved from 12 of 22 TPC-H queries to 16. The remaining six fail because the deparser cannot walk a join tree with joins on both sides, which the post says out loud instead of burying. A benchmark headline with the unfinished work listed underneath is rarer than it should be.