Power outages, hardware failures, and data center upgrades are just a few of the many problems that still bubble up to the engineering teams responsible for systems. Data center upgrades are common, and given enough time at AWS, your product team will get an email or notification stating that some servers will shut down, or experience brownouts, or small outages of power. We've shown that the best way to handle these is to span across data centers (AZs) so that, if a single location experiences issues, the systems will continue to respond. Your services should be configured in an N+1 configuration. If a single frontend is acceptable, then it should be configured for two. Spanning AZs gives us further protection from large-scale outages while keeping latency low. This allows for hiccups and brownouts, as well as an influx of traffic into the system with minimal impact to the end users.
An example of this architecture can be seen in the reference architecture for Cloud Foundry...