New versions of HDF5 support a single-writer, multiple reader (SWMR) mode that keeps files intact while appending to existing datasets.

This can be helpful for CarpetIOHDF5 since it reduces the window in time during which files would be corrupted if Cactus is terminated while writing to a file. A high-level description of the functionality is found on:

https://support.hdfgroup.org/documentation/hdf5/latest/_s_w_m_r_t_n.html

Unfortunately, even for the reader, the functionality is not fully transparent it seems (namely the reader must pass flags to H5Fopen), so this may require a "fix-files" step to make partially written files readable with standard utilities (eg VisIt). But at least the files themselves would be mostly intact.

--
Ticket URL: https://bitbucket.org/einsteintoolkit/tickets/issues/2938/use-swmr-mode-for-hdf5-output-in