I only skimmed the paper pointed to by the blog post, but one quote jumped out at me:

"Hence, in an era where disk reads caused wild variation in transaction length, allowing nondeterministic reordering was the only viable option."

In other words, when all your data is in ram (possibly distributed across a cluster of machines connected by a relatively fast network), the performance tradeoffs are very different than when your data is on rotating mechanical disks.

I don't know that I'm convinced that their approach is better, but they've certainly given me something to think about (and re-read in more detail later).