Long story short: I need to run dump(..., iterable_as_array=True) because of memory concerns. With this option, however, I lose control of how numpy.ndarray is serialized, for example. I do not need numpy.ndarray to be treated as a generic Iterable: I want, for example, to store the file name instead which I can perfectly do when iterable_as_array=False or I just use the built-in json library. iterable_as_array=True has an overwhelming priority which is hard to justify.