This guy seems a little *too* optimistic about his solution. It's a great solution, and if you want to see a good example of its use, check out the
K42 operating system. It solves a bunch of the performance problems a naive implementation as the article describes would have.
However, it's not a free lunch. Runtime marshalling and switching between interfaces? Does the author have any idea how hard that is? One tiny mistake and your entire database is full of junk data because you forgot that bit-X means Y, or your regex P had a typo. if you're going to adopt this "Continuous Change" model, you need to do an awful lot of validation testing on all the possible inputs from all the different API versions, and make
absolutely sure you get the internal state you intended. Another way to say this, a hidden feature of the discrete approach is that you know the state of your system at all points in time. You're not running a franken-process.
BTW, I'm not trying to dump on the idea, I reckon it's cool, just the article doesn't really represent it accurately.