DNM Demonstrate segfault with variable-width strings

[crusaderky]

The filter segfaults when faced with char** data.

This PR demonstrates the issue on hdf5 1.4 and shows how it's fixed in the upcoming 2.1, thanks to HDFGroup/hdf5#6178.

XREFs

• Blosc/Blosc2 segfault with variable-width strings silx-kit/hdf5plugin#364
• can_apply and set_local filter callbacks skipped for H5T_VARIABLE dtypes HDFGroup/hdf5#5942

[crusaderky]

Unsure why CI is not being triggered?

[FrancescAlted]

Unsure why CI is not being triggered?

Because there is not a proper CI set yet (we typically test this plugin manually with PyTables). Would you like to provide one?

Also, I don't see the fix in the PR (apparently it was reverted in a commit?). Also missing the PR for the hdf5-blosc2 plugin.

[crusaderky]

Unsure why CI is not being triggered?

Because there is not a proper CI set yet (we typically test this plugin manually with PyTables). Would you like to provide one?

Yes, happy to work on it.

Also, I don't see the fix in the PR (apparently it was reverted in a commit?).

Yes, I'm working on it.
I just got to a roadblock. I discovered that the underlying issue is that, for H5T_VARIABLE dtypes, libhdf5 invokes neither the can_apply nor the set_local callbacks. As a result cd_values[0:5] are uninitialised (or initialised to 0, if you are using h5py). It can be worked around by forcing typesize=1 when writing, but you'll end up with uninitialised values on disk which is really not ideal. I need to investigate libhdf5 to see if I and h5py are doing anything wrong or if we need an libhdf5 fix.

Also missing the PR for the hdf5-blosc2 plugin.

Also working on it. It's the same problem.

[crusaderky]

test_strings crashes here if you remove the workaround hack from it. blosc_set_local did not run so this is valued 0.

[crusaderky]

Opened upstream issue: HDFGroup/hdf5#5942
This PR has been rebased on top of #42.