Person retrieval aims to use images of any modality or type of the target person as query probes to retrieve and locate the target individual in the scene image gallery. However, current researches on person retrieval still focuses on the close world with controlled environments and abundant annotations, which limits its applicability to the open world in Cyber-Physical-Social Systems (CPSS). Open environments are confronted with challenges such as complex and variable scenarios, scarce labels, massive data volume, and scattered storage. To tackle these challenges, we propose a novel CPSS-based framework for open world person retrieval and its distributed parallel training. This framework comprises three tightly integrated components: a physical part that includes image data collected from various sensors; a social part composed of open world person retrieval models; and a cyber part that incorporates distributed parallel training strategies for these models. In addition, we introduce a new evaluation index that balances model performance and computational cost to assess open world person retrieval models and their distributed parallel training strategies. We hope that this work will contribute to advancing research and practical deployment of person retrieval in the open world.