Motivation of tf.data.Dataset

What would you expect as methods from a dataset type?
Map, filter, batch, pad, shuffle, repeat, persistence, iteration, from_generator, interleaving
prefetching and caching
But: no (native) train test split
close integration with keras training methods