As discussed IRL, I am a bit worried by issuing a FutureWarning whenever we call fetch_openml with the default arguments, even when there are no missing values in with old and new pandas.

We could parse the CSV twice with the old and new missing value markers to detect if its worth raising the FutureWarning. It's 2x the cost but maybe it's not that bad and off course we would not do it when the user is explicit about the missing value markers and furthermore, it's a transient overhead: after 2 scikit-learn releases we would be back to parsing the data only once (with the new default missing value markers).

Parsing twice is going to be very slow. This particular case only happens with "None" as a string. We could also check if that exists maybe and only warn in those cases?

Also, for fetch_openml, I hope there aren't many "None" values used to indicate missing value, so we could also assume it doesn't exist and set the parameter so that we use the old pandas behavior here.

Also, for fetch_openml, I hope there aren't many "None" values used to indicate missing value, so we could also assume it doesn't exist and set the parameter so that we use the old pandas behavior here.

Actually, this is an interesting point. The ARFF format forces us to use "?" (cf. https://www.cs.waikato.ac.nz/~ml/weka/arff.html) as a missing marker. I forgot to set keep_default_na_values=False when making the parser meaning that we set additional markers for missing values.

Theoretically, this could be considered a bug since we don't follow the ARFF specification.
We could simply set keep_default_na_values=False and make it a bug fix. It might introduce some regressions as well.

We could still merge #26433 to allow more flexibility if the ARFF uploaded is not following strictly the ARFF specs.

Happy to merge the other one, so we close this?

When doing the new PR, I will indicate that it will close this one.