There’s a classic tradeoff between safety and liveness in the context of replicated data systems, originally proposed by Eric Brewer and later made precise by Seth Gilbert and Nancy Lynch. The idea is that, in a networked system of servers that is vulnerable to partitions (that is, where messages betweeen servers can be arbitrarily delayed, perhaps forever), it’s impossible to ensure both that when servers respond to requests, it’s with the “right” response (consistency, a safety property), and that every request eventually receives a response (availability, a liveness property).