[JOSS review] Python version and Dependecies

[cheginit]

I am wondering what's reason behind not supporting Python 3.7+. What specific features of Python 3.9 is being used in the code? Pinning the dependencies can limit the usage of the project with other packages. Although one can create several environments, having such unnecessary pinned dependencies might affects the user experience.

The requirements.txt includes unnecessary dependencies:

recommonmark
setuptools
sphinx
sphinx-rtd-theme

Some other suggestions:

• I suggest to use conda instead of pyenv, since you're relying on scientific Python libraries, they tend to be orders of magnitudes faster when installed via conda-forge.
• You don't think you need both pyarrow and fastparquet since they both do the same thing. pyarrow is the default parquet engine in pandas.
• When using h5netcdf instead of netcdf4 you need to be aware of, and inform the user about, the difference that are documented on xarray's website:

There may be minor differences in the Dataset object returned when reading a NetCDF file with different engines. For example, single-valued attributes are returned as scalars by the default engine=netcdf4, but as arrays of size (1,) when reading with engine=h5netcdf.

Overall, I think the dependencies need a reevaluation.

EDIT: Reference to openjournals/joss-reviews#3221

[thurber]

Fair point, many folks are bringing up the Python version concerns. I briefly tried to make use of some of the new shared memory semantics of 3.9, but abandoned that effort -- honestly my main concern is to push folks to keep up to date on their software tooling, but that isn't really the job of a library. I will update to ensure support for 3.7+

[cheginit]

Yes, I understand. The main issue with not supporting the Python versions that haven't reached EOL is compatibility with other libraries. Version 3.6's EOL is in 7 months and most of the major libraries have already dropped support for it. So I think 3.7+ support is reasonable.

[thurber]

In #46, I have set the minimum python version to 3.7 and moved the dependencies related to building the documentation into the extra_require section such that they are optional. Also, removed fastparquet dependency since I'm not using it anywhere.

Wanted to respond to your other points as well:

• conda -- I think this is a good idea; I need to learn how to submit a package to conda-forge, so I will make a separate issue for this
• h5netcdf vs netcdf4 -- for all the model outputs i'm using the default netcdf4 engine; the one place h5netcdf is used is when creating a compressed file from all the static grid/model variables, which is mostly just used for the unit tests (i found this engine created slightly smaller files and read them slightly faster). It is unnecessary complexity though so I will make another issue to back out this requirement.

[cheginit]

That's looks great!

Regarding conda, that's a (strong) suggestion and not a requirement for JOSS. For development and testing, I highly recommend MambaForge. For Github Actions, you can take a look at the action file that I am using in one of my projects. It uses MambaForge and caching.

Regarding h5netcdf, I agree that it's an unnecessary complexity.