custom `numpy` dtypes

[keewis]

I wonder if it would be worth investigating if it is possible to implement a custom numpy dtype (see NEP 40-43), similar to geopandas' GeometryArray? We can probably try to interface with odc-geo for this.

There's not a lot of documentation so far (besides the NEPs), but there is a experimentation repository that can serve as a reference. It is written in C, though.

[snowman2]

I wonder if this is relevant: pydata/xarray#5102

[djhoese]

Yeah I was thinking the same thing @snowman2. I think because we aren't dealing with dataframes, we'd need to talk about coordinates or indexes of a certain numpy dtype.

[keewis]

I can't point you to a specific place, but it seems there's work in progress to use awkward to represent polylines / polygons in memory which would make this suggestion obsolete.

Regarding the index: what exactly did you have in mind? There is a temporary API to set indexes, so we could start working on that (and the more indexes we have as a reference, the easier it will be to define a good final public API).

[keewis]

see pydata/xarray#4285 for some related discussion

[keewis]

shapely 2.0 apparently will be exactly that: numpy arrays of geometries with spatial operations as numpy ufuncs (though those arrays will still have the object dtype).

See also this discussion on the pangeo forum for some discussion and a prototype of a geometry index.