In the past decade, network diagnosis has transformed from a black art into a more disciplined practice, aided by formal methods, particularly in clean-slate networks like datacenters, which are over-provisioned with well-defined control software and well-measured traffic.

But networks in the wild, from the Internet to enterprise networks and smart spaces, are mundane and still require a human user to direct the course of diagnosis to sensibly interpret an analysis or to keep the reasoning focused. One key missing is the ability to reflect, to reason not only about network states but also about the reasoning process itself. Can we, then, bring in some form of reflection as a means to more effective network diagnosis? This paper seeks an affirmative answer by showing the usefulness of reflection in explaining how a network reaches an unintended state (in addition to catching one), and in adjusting a reasoning process by exploiting knowledge of the context from which the diagnosis task arises.

We show the feasibility of reflection by an implementation in Prolog, the meta-programming support of which makes it particularly easy to construct concise meta-interpreters that serve as a reflection engine and an injection point for new diagnosis functions.