Obviously you can scale systems ahead of time if you know, but you still run into issues because you probably haven't scaled the entire system and don't know the likely bottlenecks since it's hard to generate that kind of traffic in a QA environment.

Think of the scale of facebook, and then realize most of their payloads are <1k json/protobuf. It's not the size of the payload, it's the delta in quantity.