Brent Sordyl's blog

1. Spread across as many AZs as you can.
2. Beyond mission critical? Spread across many providers.
3. Build for failure.
4. Understand your components and how they fail.
5. Try to componentize your system.
6. Test your components.
7. Relax. Your stuff is gonna break

(Full Story: How SmugMug survived the Amazonpocalypse)