Sounds correct to me. There's some chance that someone somewhere is
intentionally indexing with missing fields, and if such a person shows up
we might have to go through a deprecation cycle before making this a hard
error. I'm inclined to ignore this possibility until such a person actually
does show up, though, because if you ask for 3 fields and get 2 then you
are going to have problems in most cases already; this behavior seems
pretty difficult to (mis)use correctly.

Want to have a go at putting together a pull request?
On 5 Nov 2014 23:36, "gerritholl" notifications@github.com wrote:

If we have an ndarray with a structured dtype, and I index it with a
single non-existent field, an exception (ValueError) is raised, as
expected. However, if I index it with a list of fields, of which at least
one is non-existent, no such exception is raised:

In [402]: A = numpy.empty(shape=(5,), dtype=[("A", "f2"), ("B", "f4"), ("C", "f8")])

In [403]: A["D"] # expect: ValueError

ValueError Traceback (most recent call last)
in ()
----> 1 A["D"] # expect: ValueError

ValueError: field named D not found

In [404]: A[["C", "D"]] # expect: ValueError
Out[404]:
array([(3.822678085115078e-297,), (-3.1341142554817164e-294,),
(3.8820982892244793e-265,), (1.9848357676925587e-263,),
(1.2046357989657346e-307,)],
dtype=[('C', '<f8')])

In [405]: A[["D"]] # expect: ValueError
Out[405]:
array([(), (), (), (), ()],
dtype='{'names':[], 'formats':[], 'offsets':[], 'itemsize':14}')

I think it would be more consistent if an error would be raised in both
cases.

—
Reply to this email directly or view it on GitHub
#5262.

Fixed by #5636