It should be useful for data storage and transmission. We can always convert data during upload or download in OpenML's client APIs. For instance, people may upload a Python pandas dataframe to OpenML, and later get the same dataframe back, without realizing or caring how the data was stored in the meantime. If people want to store the data locally, they can download it in the format they like (e.g. a memory-mapped format like Arrow/Feather for fast reading or TFRecords for people using TensorFlow). Additional code can facilitate such conversions.