This paper presents a scalable local learning algorithm for large-scale classification tasks, implemented on Apache Spark. Rather than relying on a global non-linear model, our proposed approach classifies each test datapoint using a lightweight model trained on its k-nearest neighbors, assuming that local regions are less complex and more easily separable than the full training dataset. To achieve scalability, both training and test datasets are distributed across a Spark cluster, enabling parallel distance computation and independent local model training for each test datapoint. Experimental results on the ImageNet dataset, which comprises 1,281,167 images across 1,000 classes, show that the algorithm, executed on two PCs, completed the classification task in 14.7 h and achieved an accuracy of 89.58%, demonstrating its effectiveness and efficiency on large-scale data.