Abstract: The cloud downloading scheme, first proposed by us in 2011, has effectively optimized hundreds of millions of users' downloading experiences. Also, people start to build a variety of useful Internet services on top of cloud downloading. In brief, by using cloud facilities to download (and cache) the requested file from the “best-effort” Internet on behalf of the user, cloud downloading ensures the data availability and remarkably enhances the data delivery speed. Although this scheme seems simple and straightforward, designing a real-world cloud downloading system involves complicated and subtle trade-offs (between deployment cost and user experience) when serving a large number of users: 1) how to plan the cloud cache capacity to achieve a high and affordable cache hit ratio, 2) how to accelerate the data delivery from the cloud to numerous users, 3) how to handle the dense user requests for highly popular files, and 4) how to judge a potential downloading failure of the cloud. This paper addresses these design trade-offs from a practical perspective, based on big data from a nationwide commercial cloud downloading system, i.e., Tencent QQXuanfeng. Its running traces help us find reasonable design strategies and parameters, and its real performances confirm the efficacy of our design. Our study provides solid experiences and valuable heuristics for the developers of similar and relevant systems.