How to Release New Better Machine Learning Models Without Much Downtime?

So we have another question in the chats, this time from IU. And he's asking, how do you release new better models and production environments without much downtime? Quite a little production related questions. I usually speak about like starting at refactoring, because don't forget that there is this concept of exploratory operational symmetry. So it starts with refactoring edge cases, unit testing, then like a cascading of larger and larger sets,. Like 10%, 50%, 100% tests, testing in life, testing in canary, A-B testing, and then you are deploying, right? But you are using the same code through these pipes.