Test failure on 0.11.8

[limburgher]

Describe the bug
Building 0.11.8 with igraph 0.10.15:
=================================== FAILURES ===================================
_________________________ GraphTestRunner.test_labels __________________________

args = (<tests.drawing.matplotlib.test_graph.GraphTestRunner testMethod=test_labels>,)
kwds = {}

@wraps(func)
def inner(*args, **kwds):
    with self._recreate_cm():

      return func(*args, **kwds)

E matplotlib.testing.exceptions.ImageComparisonFailure: images not close (RMS 3.652):
E result_images/test_graph/graph_labels.png
E result_images/test_graph/graph_labels-expected.png
E result_images/test_graph/graph_labels-failed-diff.png

/usr/lib64/python3.13/contextlib.py:85: ImageComparisonFailure
=========================== short test summary info ============================
FAILED tests/drawing/matplotlib/test_graph.py::GraphTestRunner::test_labels
======================== 1 failed, 509 passed in 22.50s ========================

[szhorvat]

I tried this with Fedora Rawhide in Docker.

If I use the test.sh script, which installs dependencies from PyPI (matplotlib 3.9.2), all tests pass.

If I use Fedora's own matplotlib (version 3.9.1), I do see this failure. It's not due to igraph, but due to matplotlib, or more likely something matplotlib depends on for font rendering. The difference is that the label "5" is rendered one pixel lower.

Expected image:

Actual image:

Diff:

[limburgher]

Interesting.

[QuLogic]

PyPI wheels are always built with FreeType 2.6. System builds use system dependencies, with latest FreeType that renders a bit different. This would not be limited to Fedora, but also, e.g., conda-forge unless you were using their testing label (which bundles older FreeType.)

If you want to check, you can look at matplotlib.ft2font.__freetype_version__ and set a higher tolerance, or you can leave it to downstreams to patch.

[ntamas]

@szhorvat Can we just bump the tolerance then? Since you already set things up in Docker, can you suggest a new threshold that we could use?

[szhorvat]

@limburgher Can you please ignore this test for now and get this version into Fedora? The failure clearly doesn't indicate a bug. We'll find a way to deal with the test.

[limburgher]

On it!

[szhorvat]

@ntamas That's going to be a pretty huge bump in tolerance. Currently it is 0.025. If I set 3.0, it still fails. If I set 4.0, it passes. I'm experimenting here:

python-igraph/tests/drawing/matplotlib/utils.py

Lines 72 to 76 in 912f7d9

	def image_comparison(
	baseline_images,
	tol=0.025,
	remove_text=False,
	savefig_kwarg=None,

What do we do then?

Any input on this, @iosonofabio ?

[szhorvat]

3.66 is the lowest tolerance that passes. To do further experimentation, I'd have to re-create the now-deleted Docker environment, which I hope to avoid.

[ntamas]

I don't really mind bumping the tolerance to 4.