A recent line of works apply machine learning techniques to assist or rebuild cost based query optimizers in DBMS. While exhibiting superiority in some benchmarks, their deficiencies, e.g., unstable performance, high training cost, and slow model updating, stem from the inherent hardness of predicting the cost or latency of execution plans using machine learning models. In this paper, we introduce a learning to rank query optimizer, called Lero, which builds on top of the native query optimizer and continuously learns to improve query optimization. The key observation is that the relative order or rank of plans, rather than the exact cost or latency, is sufficient for query optimization. Lero employs a pairwise approach to train a classifier to compare any two plans and tell which one is better. Such a binary classification task is much easier than the regression task to predict the cost or latency, in terms of model efficiency and effectiveness. Rather than building a learned optimizer from scratch, Lero is designed to leverage decades of wisdom of databases and improve the native optimizer. With its non intrusive design, Lero can be implemented on top of any existing DBMS with minimum integration efforts. We implement Lero and demonstrate its outstanding performance using PostgreSQL. In our experiments, Lero achieves near optimal performance on several benchmarks. It reduces the execution time of the native PostgreSQL optimizer by up to 70% and other learned query optimizers by up to 37%. Meanwhile, Lero continuously learns and automatically adapts to query workloads and changes in data.