Random sampling has been widely used in approximate query processing on large databases, due to its potential to significantly reduce resource usage and response times, at the cost of a small approximation error. We consider random sampling for answering the ubiquitous class of group-by queries, which first group data according to one or more attributes, and then aggregate within each group after filtering through a predicate. The challenge with group-by queries is that a sampling method cannot focus on optimizing the quality of a single answer (e.g. the mean of selected data), but must simultaneously optimize the quality of a set of answers (one per group). We present CVOPT, a query- and data-driven sampling framework for a set of group-by queries. To evaluate the quality of a sample, CVOPT defines a metric based on the norm (e.g. $\ell_2$ or $\ell_\infty$) of the coefficients of variation (CVs) of different answers, and constructs a stratified sample that provably optimizes the metric. CVOPT can handle group-by queries on data where groups have vastly different statistical characteristics, such as frequencies, means, or variances. CVOPT jointly optimizes for multiple aggregations and multiple group-by clauses, and provides a way to prioritize specific groups or aggregates. It can be tuned to cases when partial information about a query workload is known, such as a data warehouse where queries are scheduled to run periodically. Our experimental results show that CVOPT outperforms current state-of-the-art on sample quality and estimation accuracy for group-by queries. On a set of queries on two real-world data sets, CVOPT yields relative errors that are 5x smaller than competing approaches, under the same space budget.