Pierre Augier

Yes, it is exactly the case and the issue. IMHO, it is for most packages with extensions the wrong behavior. There can be different reasons so that the same source...

Note that the name `--no-binary` is quite misleading if it actually means "Do not use manylinux wheels from PyPI, but do use binary packages previously compiled locally". If we cannot...

For the record, this works: ``` pip cache remove fluidfft_fftwmpi*; pip install fluidfft-fftwmpi --no-binary fluidfft-fftwmpi ``` Note the `_` for the first command (remove). This is a pattern for the...

Note that there are examples on the web where `--no-binary` is used in a way that just might work. For example, in https://docs.pydantic.dev/1.10/install/#performance-vs-package-size-trade-off, there is ```sh SKIP_CYTHON=1 pip install --no-binary...

Another problem related to `pip cache remove` using a pattern for the wheels filename and not the package name (as noted in https://github.com/pypa/pip/issues/12954#issuecomment-2407652674): ``` pip cache remove pyfftw; pip install...

I created four distinct and tractable issues about this problem: - doc: https://github.com/pypa/pip/issues/13087 - https://github.com/pypa/pip/issues/13086 - https://github.com/pypa/pip/issues/13090 - https://github.com/pypa/pip/issues/13088

Thanks for the explanation. It makes sense but this also leads to very bad behavior, as presented in https://github.com/h5netcdf/h5netcdf/issues/302.

There should be an API to be able to keep the old behavior (using `DATASPACE SCALAR` for numbers).

I understand your (h5netcdf) difficult situation. I'm going to open an issue on netCDF4-python. To be honest, I don't understand if there is a good reason for netCDF-C and netCDF4-python...

For context, before h5netcdf 1.7, I was using h5py to read the attributes in .nc files saved by h5netcdf and it was working fine, but I cannot do that anymore...