Scalable distributed join processing in a parallel environment requires a partitioning policy to transfer data. Online theta-joins over data streams are more computationally expensive and impose higher memory requirement in distributed data stream management systems (DDSMS) than database management systems (DBMS). The complete bipartite graph-based model can support distributed stream joins, and has the characteristics of memory-efficiency, elasticity and scalability. However, due to the instability of data stream rate and the imbalance of attribute value distribution, the online theta-joins over skewed and varied streams lead to the load imbalance of cluster. In this paper, we present a framework D-JB (Dynamic Join Biclique) for handling skewed and varied streams, enhancing the adaptability of the join model and minimizing the system cost based on the varying workloads. Our proposal includes a mixed key-based and tuple-based partitioning scheme to handle skewed data in each side of the bipartite graph-based model, a strategy for redistribution of query nodes in two sides of this model, and a migration algorithm about state consistency to support full-history joins. Experiments show that our method can effectively handle skewed and varied data streams and improve the throughput of DDSMS.