WIP: Enh appveyor debug

[yarikoptic]

No description provided.

[effigies]

Relevant external threads:

• AppVeyor: https://help.appveyor.com/discussions/problems/17301-stochastic-failure-on-python-34-and-35-builds
• Apparent is/equality failures in Python 3.4, 3.5 on Windows numpy/numpy#12096

[yarikoptic]

FWIW @effigies - numpy is the same but the dtype is a different beast:
https://ci.appveyor.com/project/nipy/nibabel/builds/19350533/job/89d24ohp5edv1uyv#L831

mnc.get_data_dtype().type  ID: 1935635128, __module__: numpy, id(sys[__module__]): 53534272
              tp['dtype']  ID: 1935634912, __module__: numpy, id(sys[__module__]): 53534272

[nibotmi]

☔ The latest upstream changes (presumably #678) made this pull request unmergeable. Please resolve the merge conflicts.

[effigies]

Just as an update, still happening intermittently on master: https://ci.appveyor.com/project/nipy/nibabel/builds/19462742

[yarikoptic]

I think our best bet is to try to bisect it to when it started to happen/commit which lead to it (if it is nibabel issue and now external library/python). When did you say it started -- June'ish?

[effigies]

June 8 was the first I could find digging through AppVeyor history. It was on #638.

I found nothing in AppVeyor's logs that indicate a change to 3.4 or 3.5 around that time. I haven't looked at numpy releases. Perhaps we should start by pinning older versions of numpy?

[yarikoptic]

so the code before #638 merge does or doesn't fail if we run it in current appveyor setup? ;-)

[effigies]

I tried reverting only #638 (see #677) and the errors persisted, but I haven't tried reverting to dfd113c.

[yarikoptic]

it might be worth ;-) and if fails - a month or two before... also - does appveyor merge master prior running or just uses PR's version?

[effigies]

This effort is dead. The bug seems to be happening less lately, so hopefully whatever was going on over at AppVeyor is resolved. I think if this recurs, we should think about moving to Azure Pipelines for Windows tests.