In my opion, Cephfs and GlusterFS have much better support for Pandas and others, you don't need change any code between local development and train an distrute

In my opion, Cephfs and GlusterFS have much better support for Pandas and others, you don't need change any code between local development and train an distruted way, rather than a path. Second, the Python library for HDFS provides only subset of filesystem API, which limits what you can do with it.