Thanks @ikrommyd,

this looks good to me!
__array__ will be triggered if anything tries to convert the awkward array to a numpy array implicitly. At that point you can't always use ak.to_numpy (because that may happen in 3rd party libraries, e.g. matplotlib maybe?), thus we have to make sure that __array__ works correctly.

Thanks @ikrommyd,

this looks good to me! __array__ will be triggered if anything tries to convert the awkward array to a numpy array implicitly. At that point you can't always use ak.to_numpy (because that may happen in 3rd party libraries, e.g. matplotlib maybe?), thus we have to make sure that __array__ works correctly.

@pfackeldey - thanks for checking this. Let’s discuss it at our next meeting. Implicit copy is something we want to watch out for or actively prevent.

I would like to also add that before this PR, we were copying more than we needed because of this:

    if dtype is None:
        return out
    else:
        return numpy.array(out, dtype=dtype)

That meant that np.asarray(some awkward array with int64 dtype already, dtype=np.int64)would be copied because numpy.array always copies and that's very unnecessary. This statement "either performs a zero-copy conversion or raises an error." in the __array__ docstring was also wrong because it would actually copy more than it should even. Numpy clearly recommends to error only if copy=False and a copy is required. copy=None should only copy if required and never error and copy=True should always copy. This is the expected __array__ behavior from numpy.

Needs tests

Hi @ianna, I have added tests for both numpy2 and numpy1 because of slight differences regarding where the copy kwarg is available and its possible values.

Anyone know what to do about the docs failure? It looks like it's timing out when reading the chicago taxi parquet dataset.

Seems like the docs building ci finally didn't crash the last time. Are we good to merge this? 😃

Seems like the docs building ci finally didn't crash the last time. Are we good to merge this? 😃

Please hold on. We’ll do it later