Sure, but rolling back work that was already done is a waste of development resources.

Disagree. It's much more wasteful to have an outage. Roll back asap, fix the issue, roll forward, do post mortem, grow as an organization. Never repeat the same mistake.

We're obviously talking past each other.

What I'm saying is that for a hosting architecture to make it difficult to predict the cost of any code change is a downside compared to an architecture that makes such predictions easy and intuitive.

Of course you will try to mitigate any downsides and learn what you can from any mistakes. But unpredictability makes learning far more difficult than it should, which inevitably means a waste of development resources.