Stateless applications aren't: they rely on other systems storing and transforming shared state correctly. On the basis of documentation and reputation we assume that our clients and database systems comprise a safe, meaningful distributed system. How justified is our faith in that system's correctness? Do popular databases actually provide the safety invariants we assume? Are we using those invariants correctly? I want experimental answers to these questions.
Kyle Kingsbury is the author of Riemann, Jepsen, Knossos, Clojure from the Ground Up, and a bunch of other open-source stuff. He writes Clojure and helps monitor distributed systems at Factual.