IIRC, it was a deliberate change. Do you mean KeyError?

I had a quick look at the recent release notes, but I could not find this change documented there. I'm using Python 2.7

Are you sure? I just checked out 1.8.0 and it gives ValueError. 1.8.3 and 1.9.3 also give ValueError.

I made a couple PRs starting in 1.10 which modified a lot of this code, but I made sure to leave it as ValueError to preserve compatibility. As you can see in #5636, for example, I wanted to change it to KeyError, however I held back because of backward compatibility. There are a number of unit tests that explicitly check for ValueError.

Interesting. Perhaps there is a bug on our end, that only manifests with the older version 1.8.

Curious. What does numpy.__version__ say? And what is the exact error message when you get the IndexError?

On numpy 1.11.0, the following code raises ValueError as expected, but with numpy 1.8.2 it raises IndexError:

import numpy as np
print np.__version__
records = np.empty(5, dtype=[('a', 'f8')] )
for record in records: record['foo']

yields

1.8.2
Traceback (most recent call last):
File "", line 1, in
IndexError: invalid index

$ python --version
Python 2.7.6

$ cat /etc/lsb-release
DISTRIB_ID=Ubuntu
DISTRIB_RELEASE=14.04
DISTRIB_CODENAME=trusty
DISTRIB_DESCRIPTION="Ubuntu 14.04.4 LTS"

Oh OK, that makes sense. What actually changed behavior is indexing a "structured scalar" rather than an array (very subtle, I know!). Indeed, in #5947 I changed it so structured scalars behave like structured arrays, so they give the same errors:

>>> arr = np.empty(5, dtype=[('a', 'f8')] )
>>> scalar = arr[0]
>>> arr['b']
ValueError: no field of name b
>>> scalar['b']  # this is what changed in 1.10
ValueError: no field of name b

I didn't think to put the change in the release notes. @charris is it worth putting a belated note in the 1.11 notes?

I think we can close this issue. That the exception is not ideal is also mentioned in gh-8519.