[mechanical_fish]

It's a hard thing to engineer, especially after the fact, and especially when you are trying to hide it behind an abstraction layer. (Which is to say: You can't expect your customers to engineer their apps with multiregion in mind, or to take it kindly when you raise rates to support additional redundant hardware and bandwidth.)

e.g. because region-to-region data transfer is not free, and trans-region latency is ugly, you can't just relaunch half your instance farm in another region and expect happiness. There are also routing issues: Internal IPs don't work across regions, elastic IPs don't transfer across regions...