Short answer: the ones that you need them to be
Longer one: it's all programmable nowadays, so what are the spaces, what are the calculations etc. it's all up to you. Several common approaches are:
1. lights in world space. Then you have to xfrom your normals into world space in a vertex shader for each vertex.
2. lights in object space. Then you have to xfrom lights into object space before rendering each object.
Each has pros and cons, as always Ditto with eye position and other things.