Source: h5py Followup-For: Bug #994617 Hi Mattia, historically hdf5 has not been entirely ABI-stable, see https://forum.hdfgroup.org/t/c-c-abi-stability-and-binary-compatibility-between-patch-versions/5312 https://forum.hdfgroup.org/t/another-abi-breakage/5503
The test was added in https://github.com/h5py/h5py/pull/867 Not entirely clear what the context of the test is (the upstream issues references in the PR seem to discuss being unable to read the libhdf5 version on Windows). But the h5py test was added in 2017, while HDF5 was still discussion ABI [in]stability in 2019. But 1.10.2 and 1.10.3 were released in 2018, after which HDF5 seems to be identifying ABI correctly. They include binary compatibility reports in their Release Notes e.g. https://portal.hdfgroup.org/display/support/HDF5%201.10.8#compatibility1108 https://portal.hdfgroup.org/display/support/HDF5%201.10.3#compatibility identifies an incompatibility between 1.10.2 and 1.10.3 HDF5 1.10.3 introduced ABI 103 But their binary incompatibilities seem to be correctly reflected in ABI sonames. The discussion in 2019 addressed 1.10.1 and 1.10.2. They considered 1.10.2 the proper stable release (in 2018), and updated the ABI for 1.10.3 as needed (also in 2018). python3-h5py-serial and python3-h5py-mpi declare their dependency on the appropriate ABI-versioned libhdf5 so you're right, that should capture compatibility satisfactorily. So should be safe to relax the runtime version test. I'll keep the spirit of it by reducing it to a minor version comparison. Drew
