Pattern sampling has emerged as an efficient paradigm for extracting knowledge from large-scale tabular datasets. By stochastically drawing patterns in proportion to a specified interestingness measure, such as frequency, these methods provide an anytime mechanism able to supply users with representative patterns. Existing approaches, however, operate only on homogeneous data structures such as sequential, numerical, or transactional data. In contrast, many real-world applications produce datasets that combine several heterogeneous types of information. In this article, we introduce the first frequency-based sampling method designed specifically for heterogeneous tabular datasets. Our method employs a multi-step decomposition of the sampling process and tackles the central challenge of precisely computing, for each data instance, the exact number of heterogeneous patterns that cover it. We provide a formal proof establishing that the resulting sample is proportional to pattern frequency. A comprehensive experimental study shows the quality of the drawn patterns using several indicators.