That's extremely vague. What kind of requests? Scale to 100M requests a day with any budget for infrastructure? Or is just choosing naive autoscaling without considering costs not ok? Are the requests evenly spread out throughout the day?

What do you mean "no room for error"? Every networked application has errors. Do you mean that the application should never throw an error? Or that errors should always be retried an infinite number of times? Or that requests should not get dropped and should be guaranteed to me handled? And how do you guarantee that? It is quite impossible to have an application that serves 100M requests a day from real users and have 0 errors or dropped packets.