Social networks are ubiquitous, with online networks garnering a large portion of Web traffic. Both online and offline, social networks structures are an interesting data source whose importance has been recognized for over a hundred years. Research on social network analysis has dealt with properties of entire networks, in addition to properties of nodes or sets of nodes. A user queries a social network in pursuit of a desired outcome, such as an expert on a specific medical condition, a set of influential people to promote a new product, or a well-balanced group of database experts to form a program committee. The user may know what the desired outcome is, and may even be able to express it in a formal query language, given the right abstract predicates to represent typical social-network measures (e.g., the importance of a node or its relevance to some keywords). However, choosing the best implementations for these predicates, as well as optimal ranking functions for the results, will often be beyond the abilities of a standard user. In fact, even an expert may experience difficulty with such a task, as the quality of solutions may depend on the precise query at hand, the user preferences, and the nature of the network. This paper suggests a novel vision of a social network database system. This system incorporates abstract predicates relevant to social networks as primitive building blocks in the query language, and uses machine learning, as an integral part of the query processor, to select and improve upon the predicate implementations. The paper discusses the main features of such a system, as well as the implementation challenges.