Estimating drug-target binding affinity (DTA) is crucial for various tasks, including drug design, drug repurposing, and lead optimization. Advanced works adopt machine learning techniques, especially deep learning, to DTA estimation by utilizing the existing assay data. These powerful techniques make it possible to screen a massive amount of potential drugs with limited computation cost. However, a typical DNN-based training paradigm directly minimizes the distances between the estimated scores and the ground truths, suffering from the issue of data inconsistency. The data inconsistency caused by various measurements, e.g., Kd, Ki, and IC50, as well as experimental conditions, e.g., reactant concentration and temperature, severely hinders the effective utilization of existing data, thus deteriorating the performance of DTA prediction. We propose a novel paradigm for effective training on hybrid DTA data to alleviate the data inconsistency issue. Since the ranking orders of the affinity scores with respect to measurements and experimental batches are more consistent, we adopt a pairwise paradigm to enable the DNNs to learn from ranking orders instead. We expect this paradigm can effectively blend datasets with various measurements and experimental batches to achieve better performances. For the sake of verifying the proposed paradigm, we compare it with the previous paradigm for various model backbones on multiple DTA datasets. The experimental results demonstrate the superior performance of our proposed paradigm. The ablation studies also show the effectiveness of the design of the proposed training paradigm.