We currently store the results of each HTTP request performed by fetch_openml. It still can take a substantial amount of time to decode the fetched ARFF data into numpy arrays or sparse matrices. We could instead (or also) cache the decoded data on disk, so that the ARFF does not need to be parsed in repeated calls.

Alternatively, we can hope that OpenML provides a more compact data representation some time soon, as per openml/OpenML#388.