Only few pattern mining methods are dedicated to distributed databases. In fact, the cen- tralization of data is often less expensive than the communication of all mined patterns. To cir- cumvent this difficulty, this paper follows a parsimonious approach by sampling patterns. We propose the algorithm DDSAMPLING that draws a pattern from a distributed database propor- tionally to its interest. We demonstrate its accuracy and analyze its complexity. Experiments show on several datasets its robustness against the failures of a site or the network.