The big data from online social media represents the relationship between the actors' self-organization. It contains multi-level social entity relationship. As an emerging field in recent years, online social media sampling method has important research value and practical significance in social computing. However, there are some problems in existing methods. For example, large Markov chain is difficult to parallelize, sampling is easy to be trapped in local, and there is concerns with Markov chain burn-in process. To address those issues, the paper presents a multi stage cluster sampling for online social media big data (OSM-MSCS). The proposed method first decomposes integral cluster into small cohesive subgroups, then uses delay rejection (DR) to sample typical online social relationship with parallel processing, and finally uses Gibbs sampling methods to choose interaction relationship in different cohesive subgroups to obtain the random sequence. Experimental results show that OSM-MSCS is an effective method for online social media big data, and its sampling technique is better than BFS and MHRW.