Abstract: Log-structured merge tree (LSM-tree) is widely used as the core data structure of modern key-value stores due to its ability to utilize sequential access performance of external storage. However, it suffers from expensive merge operations to maintain the layered and ordered data organization, which induces significant write amplification. Recent researches have proposed various optimizations to mitigate write amplification, but at the cost of query performance and space utilization. A new architecture for LSM-tree based key-value store is proposed. The key idea is to leverage the key-value separation design to mitigate merge overhead, while maintaining a certain degree of order for values by using a new tree structure called vTree to improve the range query performance. In the meantime, corresponding data merge and space reclaim algorithms were developed for the vTree. The experimental results show that the key-value store developed with this architecture has well balanced performance on write, point lookup, and range query, with low space cost.