Closed 0 closed 1 year ago
Looks promising. I'm not sure why Windows seems to be a special case here.
Also run the formatter. See contrib/format.
Reformatted.
I think the failure is caused by a newer version of libhdf5. The three Windows checks that fail all report "libhdf5 v1.12.2", but the others have lower versions: v1.10.4 and v1.12.1. (Confusingly, the ones that claim to be using v1.12.1 seem to get it through HDF5_jll v1.12.2+2, but that's probably a separate issue.) I can reproduce the failure locally by using the system version of libhdf5, which is newer than v1.12.1.
I may have found the relevant HDF5 commit, which was added between v1.12.1 and v1.12.2, and appears to change the logic for determining what constitutes a partial allocation. The failing test might be checking something that's too internal to libhdf5. Is it worth keeping it?
The JLL versions do not necessary correspond to their upstream versions. See https://github.com/JuliaPackaging/Yggdrasil/blob/86f01f3d8b89c113347aec7e51f8105cd5293be1/H/HDF5/build_tarballs.jl#L9-L30 .
Could you modify the test so that it is dependent on the HDF5 C library version?
Modified failing test to be version-dependent.
Thanks
The extent is not guaranteed to be a multiple of chunk size. When partial chunks are present, the number of chunks should be rounded up in order to include them.