issue_comments: 700670814

You can see an example of using xarray with structured curvilinear coordinates here: - http://xarray.pydata.org/en/stable/examples/multidimensional-coords.html - http://xarray.pydata.org/en/stable/examples/ROMS_ocean_model.html

And with unstructured data here: - http://gallery.pangeo.io/repos/rsignell-usgs/esip-gallery/02_National_Water_Model.html - http://gallery.pangeo.io/repos/rsignell-usgs/esip-gallery/01_hurricane_ike_water_levels.html

The key concept is that xarray supports both dimensions coordinates and non-dimension coordinates. The dimension coordinates must be 1D, but the non-dimension coordinates can have any dimensionality. For a regular lat-lon grid, a variable might have dimensions like this temp(time, depth, lat, lon) A structured curvilinear 2D grid might instead look like temp(time, depth, j, i) with additional coordinate variables lon(j, i) lat(j, i) which can be used for visualization (but not, currently, for indexing) A fully unstructured mesh in 2D would instead look like temp(time, depth, cell_id) lon(cell_id) lat(cell_id)

This is exactly what netCDF does to encode these data types. Anything that can go into a netCDF file can be represented in Xarray. You just don't get the full functionality in terms of label-based selection. That will hopefully change as we implement more flexible indexing (see https://github.com/pydata/xarray/projects/1).

Another limitation of xarray is that it has no explicit notion of "cell bounds," other than recognizing these as coordinates (see #2844). Our xgcm package works around this limitation in some simple ways.