With the increased demands for wireless spectrum, dynamic spectrum sharing is emerging as an important and powerful concept. Most research in this domain is being conducted in design of cognitive radios and on specific PHY and MAC layer challenges associated with them. However, for a dynamic spectrum sharing architecture to be viable, research is needed to resolve many other challenges, e.g., in the context of real-time spectrum management and enforcement. This paper is the first to present a study of some such important architectural considerations, driven by our ongoing design and implementation of a spectrum sharing system, called Spark. We propose some promising approaches to address these challenges, and enumerate the need and opportunities for significant future research in this domain.