I do notice some downsides to using the configparser, e.g. you can't see which values are explicitly set or which are default (if the explicitly set value matches the default value). It also seems to complicate the code (e.g. having to create a virtual file with FAKE_SECTION). Considering our configurations haven't gotten any more complicated for years, I think we should be safe to assume the simple field = value format will last us a while longer, and it might be more easily readable to just write a custom parser for it. It should not take more than a few lines of codes, e.g.

set_fields ={}
with open(file, 'r') as fh:
    for line in fh.readlines():
        field, value = line.split('=', 1)
        set_fields[field.strip()] = value.strip()
config = {**_defaults, **set_fields}

Not entirely sure if I want to do that right now (I'd like this CLI done before my leave next week, and while simple I feel it should have extensive testing), but do you think we should consider that?

Remarks while working on this:

• max_retries is contained in _defaults but seems like it should be set by developers only. After discussing we probably want to avoid having such a number at all as long as we employ a decent retry policy (that waits increasingly long times).
• the configparser note from above.
• verbosity is currently not used in openml-python, I think this is a bug introduced by Rework local openml directory #987.