Automatic prostate segmentation is an important step in computer-aided diagnosis of prostate cancer and treatment planning. Existing methods of prostate segmentation are based on deep learning models which have a large size and lack of transparency which is essential for physicians. In this paper, a new data-driven 3D prostate segmentation method on MRI is proposed, named PSHop. Different from deep learning based methods, the core methodology of PSHop is a feed-forward encoder-decoder system based on successive subspace learning (SSL). It consists of two modules: 1) encoder: fine to coarse unsupervised representation learning with cascaded VoxelHop units, 2) decoder: coarse to fine segmentation prediction with voxel-wise classification and local refinement. Experiments are conducted on the publicly available ISBI-2013 dataset, as well as on a larger private one. Experimental analysis shows that our proposed PSHop is effective, robust and lightweight in the tasks of prostate gland and zonal segmentation, achieving a Dice Similarity Coefficient (DSC) of 0.873 for the gland segmentation task. PSHop achieves a competitive performance comparatively to other deep learning methods, while keeping the model size and inference complexity an order of magnitude smaller.