Modern scientific research is increasingly conducted by virtual communities of scientists distributed around the world. The data volumes created by these communities are extremely large, and growing rapidly. The management of the resulting highly distributed, virtual data systems is a complex task, characterized by a number of formidable technical challenges, many of which are of a software engineering nature. In this paper we describe our experience over the past seven years in constructing and deploying OODT, a software framework that supports large, distributed, virtual scientific communities. We outline the key software engineering challenges that we faced, and addressed, along the way. We argue that a major contributor to the success of OODT was its explicit focus on software architecture. We describe several large-scale, real-world deployments of OODT, and the manner in which OODT helped us to address the domain-specific challenges induced by each deployment.