I think it has long been recognized that there are practical boundaries to transactional capabilities - but bad engineers ignored that (to their peril) and while systems were tightly coupled the failure modes were rare or even impossible.

When you build world-scale systems, certain assumptions that were always flawed, begin to show their cracks.

To my mind, there is a sort of philosophy here as well of "better prevention of problems than detection of problems". Transactions prevent (or minimize) certain problems. But they don't scale. Detection is all that is left.