“Dropping requests is preferable to failing uncontrollably.”
Imagine a system that is exposed to uncontrollable user input (for example, a site on the internet). Any deployment will be finite in both its processing capability and its buffering capacity, and if user input exceeds the former for long enough, the latter will be used up and something will need to fail. If this is not foreseen explicitly, then an automatic out-of-memory killer will make a decision that is likely to be less satisfactory than a planned load-shedding mechanism—and shedding load means dropping requests.
This is more of a philosophy than an implementation pattern. Network protocols, operating systems, programming platforms, and libraries will all drop packets, messages, or requests when overloaded; they do so in order to protect the system so that it can recover when load decreases. In the same spirit, authors of Reactive systems need to be comfortable with the notion of sometimes deliberately losing messages.