Product lines need decision models that guide the derivation of product variants satisfying specific requirements. In dynamic product lines, whose requirements vary during runtime, these decision models are also required to support automatic product reconfigurations in response to changing requirements. However, because of the combinatorial explosion of variants, such automatic decision making suffers from poor performance with many variants. In this paper, we introduce a mathematical formulation of this scalability problem. Based on the formulation, we discuss the limitations of existing approaches to support variants scalability. These limitations are addressed by our modular approach whose decision model combines (1) the use of utility functions, and (2) various optimisations of the search space. The analysis is supported by experience with the QuA and MADAM middleware, which apply this approach to supports self-adaptation based on dynamic product lines.