Flaky test `tests/test_filter_pipeline_v2.py::test_hdf5_filters[fletcher32_True-shuffle_True-compression_9-chunk_(10, 10)]`

[valeriupredoi]

Second time we see this, in a fairly large number of trials (30 odd) but there is definitely something flaky in it: https://github.com/NCAS-CMS/pyfive/actions/runs/18843709458/job/53762175768

@kmuehlbauer would you have a minute look into it please, mate? 🍺

[kmuehlbauer]

@valeriupredoi First, we could mark that test as flaky until we know what's going on.

Can we observe, if this is always the same test combo?

FAILED tests/test_filter_pipeline_v2.py::test_hdf5_filters[fletcher32_True-shuffle_True-compression_9-chunk_(10, 10)] - ValueError: fletcher32 checksum invalid

compression = 9
chunk_size = (10, 10)
shuffle = True
fletcher32 = True

There might be issues with the random generated data, that for some reason with certain chunk settings and shuffling the filter is not applied. Maybe we need to rely on some salted random data.

We could add an h5py read part before, to check if that chain is working.

[valeriupredoi]

I was just about to comment that I am not 100% sure it fails only at that specific used case - let me mark it flaky for now, then we can look at it closer 🍺

[kmuehlbauer]

@valeriupredoi I think #141 doesn't work as intended. I was under the impression, that @pytest.mark.flaky just works without any other settings. But this special flaky mark isn't part of pytest. We would need additional package pytest-rerunfailures with some adaptions.

Or we mark the test as xfail.

[valeriupredoi]

@kmuehlbauer that's a very good point! I always thought that the flaky marker came from pytest, let me add the pytest-rerunfailures plugin to our deps - good catch, mate 🍺

[bmaranville]

Do you know if this is a pyfive issue or an h5py issue? Do the files that are failing the pyfive test pass a read test in h5py?

[kmuehlbauer]

@bmaranville Not yet, I was thinking about adding a h5py test to get behind this. Maybe we should really add this.