Re-ranking has become an indispensable part of a recommender system. Based on several online experiments, we have found that there are several complex factors affecting the performance of re-ranking, such as sequence order, the dependency between documents or the diversity of categories. The combinations and interactions of these factors lead to an extremely large search space for the re-ranking model to find an optimal solution. The online behaviour of users is very different from the data collected offline, and such a significant difference implies that the recommendation model should be trained with online interactions to avoid misleading from offline data. Therefore, it is critical to propose an optimized solution within the limited response time for a better user experience. In this paper, we propose a hierarchical cascade Top-k recommendation framework, named WallEve, which can be used to balance the accuracy and efficiency of re-ranking. WallEve employs a contextual bandit algorithm at a high level for proper diversity control, reduction in space exploration and fast convergence. A reinforcement learning model is also employed at a low level to fine-tune the permutation results. Comprehensive experiments have been conducted, and the results show that WallEve outperforms DQN with a superior performance in terms of accuracy and efficiency. The A/B test results in the Taobao online environment show that WallEve improves the UV (Unique Visitor) by 5.5%-77% in different scenarios.