Bug#1103004: skimage test failures

[Stefan van der Walt]

Hi Rebecca,

Yes, we have been trying to resolve this issue, but unfortunately it is related to internal Fortran state inside of SciPy's ARPACK implementation.

The problem is annoying: it means that successive normalized cut segmentations do not give results that are bit-for-bit the same. However, since the resulting segmentations are all correct (as much as a segmentation can be), it should not be highly impactful.

I'm sorry that I don't have a better answer for now. I think easiest would be for Debian to temporarily disable that test until we can come up with a systematic fix. Either we work around it in skimage, or we wait for a new SciPy release which replaces ARPACK, where the problem occurs.

Ongoing conversation at: https://github.com/scikit-image/scikit-image/pull/7684

Best regards,
Stéfan

On Mon, Apr 21, 2025, at 13:16, Rebecca N. Palmer wrote:
> This has failed twice on reproducible-builds, and started failing 
> between 2025-02-23 and 2025-04-20:
> https://tests.reproducible-builds.org/debian/rb-pkg/unstable/arm64/skimage.html
> No other architectures have been tried since 2025-03-26, so we don't 
> actually know whether it's arm64-specific.
>
> The same test passes in the autopkgtest.