We derive the specification of a primary partition and a partitionable fail-aware node membership service in a top-down fashion. The derived specifications are fail-aware in the sense that each client of a membership server can learn if the server currently provides its standard semantics or an exception semantics because too many failures have occurred. We first propose the specification of an ideal membership service and then transform this ideal specification step by step to derive the two fail-aware specifications that are implementable in timed asynchronous systems. In each step we address an implementation problem or a change in the system/failure model.