This ticket proposes two separate but overlapping ideas.

1. When trying to write/read Zarr files, throw an informative error if the user does not have zarr installed

just as Xarray does for other optional dependencies like Matplotlib.

import xarray as xr


da = xr.DataArray([1, 2, 3],
                  dims='x',
                  coords={'x': [10, 20, 30]}
                  )
da.plot() # Raises a more informative ImportError
da.to_zarr('./test.zarr') # Does not raise an informative ImportError

2. Create a utility function to handle optional dependency imports across the Xarray code base.

Doing git grep "raise ImportError" or git grep "raise ModuleNotFoundError" reveals that currently there is a piecemeal approach to handling optional dependencies. It would probably be helpful to have a dedicated function to handle soft dependencies across the code base, and I would be happy to add this If I am going to be adding a PR for point 1 anyways.

Describe the solution you'd like

I work on an applied neuroscience package named MNE-Python, where we have a lot of optional dependencies that are only needed for specific MNE functionalities. So borrowing from MNE, I think something like this would be nice:

# xarray.util.check.py

import importlib


def _soft_import(name, reason, strict=True):
    """Import optional dependencies, providing informative errors on failure."""
    _INSTALL_MAPPING = dict(nc_time_axis="nc-time-axis")
    package_name = _INSTALL_MAPPING.get(name, name)
    try:
        return importlib.import_module(name)
    except (ImportError, ModuleNotFoundError):
        if strict:
            raise ImportError(
                f"for {reason}, the {package_name} package is required. "
                f"Please install it via pip or conda."
            )
        else:
            return None

Then, for example, here, we can instead do:

zarr = _soft_import("zarr", reason="writing zarr files")

And if we don't have zarr installed we'd get:

ImportError: for writing zarr files, the zarr package is required. Please install it via pip or conda.

Describe alternatives you've considered

Open to any ideas 🙂

Additional context

What do folks think about this idea? Is a PR welcome?