I will begin by presenting a structural description of the kind of empirical data which arises in physical experiments, using the language of sheaf-theory. A potential feature of such data is contextuality (including as a special case non-locality) which can be neatly characterised in this setting. The feature does not arise in classical data, and a number of recent results indicate that it is the key feature in enabling quantum advantages or speedups over classical behaviour in a variety of computational scenarios. The sheaf-theoretic perspective exposes a qualitative hierarchy of strengths of contextuality and I will also mention recent efforts to quantify contextuality via linear programming. Finally I will report on work in progress on developing a resource theory for contextuality, beginning with an algebra of empirical models under whose operations contextuality is monotonically decreasing.