sunpy 5.1.1 + pytest 8 on armel & mips64el error

[mr-c]

Describe the bug

Building sunpy 5.1.1 with Pytest 8.1.1 + #7429 we still get two errors:

On armel (arm32v5): test_map_arithmetic_operations_raise_exceptions[map-warn_context0] fails, DID NOT WARN. No warnings of type (<class 'RuntimeWarning'>,) were emitted.

https://buildd.debian.org/status/fetch.php?pkg=sunpy&arch=armel&ver=5.1.1-4&stamp=1712238082&raw=0

On mips64el: [doctest] sunpy.coordinates.frames.Helioprojective.assume_spherical_screen:

Differences (unified diff with -expected +actual):
    @@ -2,4 +2,4 @@
         [(   0., 0., 0.99660825), ( 319., 0., 0.99687244),
          ( 638., 0., 0.99778472), ( 957., 0., 1.00103285),
    -     (1276., 0., 1.00125872), (1595., 0., 1.00125872),
    -     (1914., 0., 1.00125872)]>
    +     (1276., 0.,        nan), (1595., 0.,        nan),
    +     (1914., 0.,        nan)]>

https://buildd.debian.org/status/fetch.php?pkg=sunpy&arch=mips64el&ver=5.1.1-4&stamp=1712238826&raw=0

Any guidance on how we shall fix this?

To Reproduce

Here's a dockerfile, build it with --platform=linux/arm/v5 or --platform=linux/mips64le

FROM docker.io/debian:sid-slim

WORKDIR /build

RUN apt-get update && apt-get install -y --no-install-recommends devscripts curl debian-keyring
RUN dget https://deb.debian.org/debian/pool/main/s/sunpy/sunpy_5.1.1-4.dsc
WORKDIR /build/sunpy-5.1.1
RUN apt-get -y build-dep .
RUN dpkg-buildpackage -uc -us -b

For reference, all the extra patches we apply to Sunpy 5.1.1 are at https://salsa.debian.org/debian-astro-team/sunpy/-/tree/master/debian/patches?ref_type=heads

Screenshots

No response

System Details

==============================
sunpy Installation Information
==============================

General
#######
OS: Debian GNU/Linux (trixie/sid, Linux 6.1.0-18-amd64)
Arch: 64bit, ()
sunpy: 5.1.1
Installation path: sunpy.egg-info

Required Dependencies
#####################
astropy: 6.0.1
numpy: 1.26.4
packaging: 24.0
parfive: 2.0.2

Optional Dependencies
#####################
asdf: 3.1.0
asdf-astropy: Missing asdf-astropy>=0.1.1; extra == "asdf" or "docs" or "tests"
beautifulsoup4: 4.12.3
cdflib: Missing cdflib!=0.4.0,!=1.0.0,>=0.3.20; extra == "docs" or "tests" or "timeseries"
dask: 2023.12.1+dfsg
drms: 0.7.1
glymur: Missing glymur!=0.9.5,>=0.9.1; extra == "docs" or "jpeg2000" or "tests"
h5netcdf: 1.3.0
h5py: 3.10.0
lxml: 5.2.1
matplotlib: 3.6.3
mpl-animators: 1.1.1
pandas: 2.1.4+dfsg
python-dateutil: 2.9.0
reproject: 0.13.0
scikit-image: 0.22.0
scipy: 1.11.4
sqlalchemy: 1.4.50
tqdm: 0.0.0
zeep: 4.2.1

Installation method

from source (.tar.gz)

[Cadair]

Thanks for the bug report (and testing sunpy on other platforms) we will try and take a look.

[ayshih]

On mips64el: [doctest] sunpy.coordinates.frames.Helioprojective.assume_spherical_screen:

Differences (unified diff with -expected +actual):
    @@ -2,4 +2,4 @@
         [(   0., 0., 0.99660825), ( 319., 0., 0.99687244),
          ( 638., 0., 0.99778472), ( 957., 0., 1.00103285),
    -     (1276., 0., 1.00125872), (1595., 0., 1.00125872),
    -     (1914., 0., 1.00125872)]>
    +     (1276., 0.,        nan), (1595., 0.,        nan),
    +     (1914., 0.,        nan)]>

https://buildd.debian.org/status/fetch.php?pkg=sunpy&arch=mips64el&ver=5.1.1-4&stamp=1712238826&raw=0

Apparently, on mips64el, numpy.fmin() does not behave as it should, namely ignoring NaNs (see the long discussion in numpy/numpy#23158). I presume that no fix was discovered for the NumPy build on debian (otherwise we wouldn't be having the problem here). We could refactor code to avoid the problematic call in our code:

sunpy/sunpy/coordinates/frames.py

Line 581 in 386e540

d = np.fmin(d, dd) if self._spherical_screen['only_off_disk'] else dd

since in our specific case, the NaNs are from taking the square root of a negative number.

[nabobalis]

This looks to be a bug in Python, which was only fixed in python/cpython#104202 which came out in 3.12.

[Cadair]

I have to say I don't think it would be worth us working around this for an architecture I very much doubt anyone is using SunPy on 😁

[ayshih]

On armel (arm32v5): test_map_arithmetic_operations_raise_exceptions[map-warn_context0] fails, DID NOT WARN. No warnings of type (<class 'RuntimeWarning'>,) were emitted.

https://buildd.debian.org/status/fetch.php?pkg=sunpy&arch=armel&ver=5.1.1-4&stamp=1712238082&raw=0

It looks like dividing by zero on ARM (with software floating point?) might not always raise a NumPy warning. See scikit-image/scikit-image#3335 for a similar-sounding issue. We can easily change the warning catching to be a warning squashing (using numpy.errstate()) since the presence of a NumPy warning is not relevant to the unit test.

[ayshih]

Upon further poking, the fact that we are even dividing in that situation is a bug, so #7551 stops that from happening

[mr-c]

Apparently, on mips64el, numpy.fmin() does not behave as it should, namely ignoring NaNs (see the long discussion in numpy/numpy#23158). I presume that no fix was discovered for the NumPy build on debian (otherwise we wouldn't be having the problem here).

Indeed, for the silx package in Debian, we skip the problematic test on mips64el https://salsa.debian.org/science-team/silx/-/blob/master/debian/patches/0009-Skip-testing-nanmin-on-float64-arrays-on-mips-el.patch?ref_type=heads

Shall I do the same for sunpy and the sunpy.coordinates.frames.Helioprojective.assume_spherical_screen doctest?

[ayshih]

This looks to be a bug in Python, which was only fixed in python/cpython#104202 which came out in 3.12.

That CPython PR changed things, but evidently did not resolve this issue because the debian log shared above has the test failing under Python 3.12.2.

Shall I do the same for sunpy and the sunpy.coordinates.frames.Helioprojective.assume_spherical_screen doctest?

After discussion among the development team, our preference is to simply skip this doctest. Modifying our code would be clunky and/or affect performance for non-mips64el platforms.

[mr-c]

Thanks, @ayshih ! Oddly enough, the mips64el issue went away when I rebuilt with #7551 cherry-picked on top

https://buildd.debian.org/status/fetch.php?pkg=sunpy&arch=mips64el&ver=5.1.1-5&stamp=1712345380&raw=0

[ayshih]

Odd indeed!

[Cadair]

Thanks @mr-c