Replace servers instead of repairing them. What that means for a team with ten machines, not ten thousand.
The idea is simple: once a server is running, you do not change it. To change something, you build a new image, test it and replace the old machine. The old one is never patched by hand.
What you gain
Servers stop drifting. Two machines built from the same image behave the same way, and a rollback is a matter of starting the previous image.
What it costs
You need a repeatable build, somewhere to keep images, and state that lives outside the machine. Databases and uploaded files cannot be thrown away with the server.
Start small
Pick one stateless service. Script its build, deploy it by replacing instances, and keep the old image for a week. If that goes well, move the next service. There is no need to convert everything at once.



