We live in an era in which scientific discovery is increasingly driven by data exploration of massive datasets. Scientists today are envisioning diverse data analyses and computations that scale from the desktop to supercomputers, yet often have difficulty designing and constructing software architectures to accommodate the heterogeneous and often inconsistent data at scale. Moreover, scientific data and computational resource needs can vary widely over time. The needs grow as the science collaboration broadens or as additional data is accumulated; the computational demand can have large transients in response to seasonal field campaigns or new instrumentation breakthroughs. Cloud computing can offer a scalable, economic, on-demand model that is well matched to some of these evolving science needs. This paper presents two of our experiences over the last year — the Terapixel Project, using workflow, high-performance computing and non-structured query language data processing to render the largest astronomical image for the WorldWide Telescope, and MODISAzure, a science pipeline for image processing, deployed using the Azure Cloud infrastructure.