Global positioning system (GPS) receivers are widely deployed in navigation and tracking devices installed in taxis, buses, utility vehicles, and smart phones. These GPS receivers can provide probe data that illuminates automotive traffic conditions for applications including traffic analysis, travel time estimation, map building, and congestion and accident detection. The characteristics of real GPS probe data are not well understood, however, and these characteristics may pose difficulties for existing algorithms. For example, map matching is often a prerequisite for GPS trace-related opplications. Existing solutions often rely on assumptions, e.g., Gaussian GPS noise, that are not necessarily true of data collected on metropolitan streets. In this paper, we study probe data from tens of thousands of taxis in the three largest cities in China. We comprehensively characterize the data, quantify noise present in it, gauge the applicability of existing map matching algorithms to the data, identify challenges in processing large-scale GPS data, and suggest research opportunities.