Would recommend a recursive strategy:

def _has_metadata(dt):
    if dt.metadata is not None:
        return True
    elif dt.names is not None:
        return any(_has_metadata(dt[k]) for k in dt.names)
    elif dt.subdtype is not None:
        return _has_metadata(dt.base)
    else:
        return False

My main question is whether we should give a warning, or rather just raise right away. The error/warning message could possibly include either use pickle directly or use `arr.view` to remove the metadata?

Erics recursive version looks good to me.

use arr.view to remove the metadata

Possibly with a comment saying "good luck building a suitable dtype from your nested structured type while preserving field order and padding"...

I think the only viable options available to users are:

• Suppress the warning (perhaps we want a subclass of UserWarning to make this easy?) - if the resulting file is unreadable rather than just not preserving metadata, this isn't a viable choice either.
• Eliminate the metadata at the source
• Use pickle

Fair point, using view is only viable for unstructured dtypes... I was toying with the idea of adding dtype.drop_metadata() that is recursive, but I do not really want to rush that in at this time.

Suppress the warning (perhaps we want a subclass of UserWarning to make this easy?)

This seems like the best option for 1.18, adding a subclass MetadataWarning

Closing and reopen, to re-trigger failing tests.

@seberg Ping.

Thanks Matti and Eric!