In this talk, Bin Fan shares his insights on data access challenges in ML applications, with particular emphasis on how Alluxio's distributed caching helps bridge the gap between storage and compute in preprocessing, pretraining and inference.