Person re-identification via 3D skeletons is an important emerging research area that triggers great interest in the pattern recognition community. With distinctive advantages for many application scenarios, a great diversity of 3D skeleton based person re-identification (SRID) methods have been proposed in recent years, effectively addressing prominent problems in skeleton modeling and feature learning. Despite recent advances, to the best of our knowledge, little effort has been made to comprehensively summarize these studies and their challenges. In this paper, we attempt to fill this gap by providing a systematic survey on current SRID approaches, model designs, challenges, and future directions. Specifically, we first formulate the SRID problem, and propose a taxonomy of SRID research with a summary of benchmark datasets, commonly-used model architectures, and an analytical review of different methods' characteristics. Then, we elaborate on the design principles of SRID models from multiple aspects to offer key insights for model improvement. Finally, we identify critical challenges confronting current studies and discuss several promising directions for future research of SRID.