In the next decade, networks of devices will be widely deployed for measurement, detection and surveillance applications. Millions of sensors and small-scale mobile devices will integrate processors, memory and communication capabilities. Large collections of devices need to be controlled and accessed in an ad-hoc manner. This paper shows that database technology can be adapted to meet the challenges of this new computing environment. In our new concept of a device database system, individual devices are modeled as database objects, which allows us to access collections of devices with declarative queries. We present a semantics for queries over device database systems and novel query processing techniques for evaluating such queries. We describe our implementation of these techniques in the Cornell COUGAR system and present a experimental evaluation that illustrates the performance characteristics of query processing in a device database system.