Here is how they found the silver bullet at Google: https://medium.freecodecamp.com/how-google-builds-a-web-framework-5eeddd691dea#.3qxfhdvoq
You can change the actual in-use version of any code, as long as you run the test scripts for everything that depends on it. This prevents dependency problems, while still allowing the system to adapt. It fixes the large system problem. I think it really is that simple… if you stay inside Google.
The one weird trick is that they have all of the dependent code in one repo, so they can easily go fix it. It gets more complicated if you are publishing a library or a service that people are using at other companies. You usually can’t go into their code and fix it to work with your change. However, you could set up a testbed where you run their tests, so that you at least know when the change causes problems, and you can fix the problems. I think that Web service ecosystems will need these testbeds.