Kai Mühlbauer

Reporting issues is a good idea anyway, so thanks for taking time reporting this. Looking forward to anything you can provide to narrow this down. I'll try searching through the...

@aidanheerdegen Did you come any further with that issue? I would try to look into this if you could provide more information, MCVE would be nice.

Thanks for coming back to this. Would you be able to provide a test UM file which broke with your setup?

I've come up with a workaround which also has the nice advantage to be a bit more performant when creating millions of attributes ;-) https://github.com/h5py/h5py/pull/2042 FYI: @paugier

I've tested that patch a bit. So the `track_order` problem is avoided then. But a major pain point is, that every change of an attribute would delete and recreate it....

Yes, netCDF4 suffers from the same issue. At least first tests suggest this.

```python import netCDF4 with netCDF4.Dataset("test.nc", mode="w") as nc: g = nc.createGroup("grp0") g.test12 = 12 g.test0 = 0 g.test1 = 1 g.test2 = 2 g.test3 = 3 g.test4 = 4 g.test5...

@hmaarrfk So, only if the data type is the same, netCDF4 just modifies the entry. `h5py` recreates in any case atm. So we would need to fix either in `h5py`...

> But I'm wondering why `test9` is listed after `test1` although it has been changed the other way round. Further tests indicate, that this might be connected with [attr_phase_change](https://api.h5py.org/h5p.html?highlight=phase_change#h5py.h5p.PropOCID.set_attr_phase_change) setting....

`netcdf-c` deletes the attribute in the hdf5 file (`H5Adelete`) when renaming. It keeps the renamed attribute in internal netcdf-c cache, so it get's written to the file in consequence. The...