issue_comments

As I was dissatisfied with the prototype, I scrapped it and rewrote it mocking the splrep/splev API. However my functions don't wrap around scipy.interpolate.splrep/splev, as those don't accept an n-dimensional y, but instead they wrap around scipy.interpolate.make_interp_spline and scipy.interpolate.BSpline (which is what scipy.interpolate.interp1d does too). Compared to the prototype above:

• lost support for Akima, PCHIP, and the non-spline options of interp1d
• MUCH more memory-efficient than before, particularly on distributed
• no more hacks - splrep produces a plain Dataset, which can be stored on NetCDF, sliced, etc. etc.
• gained ability to have chunks on x_new

I built a production-quality version (inclusive of documentation, unit tests, and all the trimmings) at https://github.com/crusaderky/xarray_extras. Happy to discuss moving it to somebody else's module.

You still can't have a chunked x. It is possible to implement it with dask.array.ghost.ghost, although it would be mutually exclusive with a chunked x_new - contributions are welcome.

Closing this ticket as I agree this is beyond the scope of the core xarray package.

This functionality sounds similar to reindex. Is it confusing to add this to reindex with an additional method such as method='linear'?

I think this would be a little confusing because then reindex in xarray would work differently from reindex in pandas.

Also, interpolate is a more descriptive name than reindex, at least for scientists. I would rather read array.interpolate_at(x=points) than array.reindex_like(x=points, method='linear').

I am very much in favor of an xarray built-in interpolate_at method. Interpolating/regridding is domain agnostic and would be extremely useful to many downstream packages (salem does some regridding but it all happens on memory and uses scipy's RegularGridInterpolator under the hood). I would love to have an xarray func for this...

For my use case slrep caching is critical, as I need to interpolate 20-something curves roughly 4000 times on different points. Changing the application to gather all points from downstream and do one big interpolation would not be feasible as it would kill off my RAM and be very hostile to a distributed environment.

Regarding interpolation, this is one of the most common needs of users. I would support adding some sort of basic interpolation directly within xarray. I would eventually like to see multilinear n-dimensional interpolation supported.

I personally think that it would be nicer if we merge these similar packages and make them more complete.

xrft is focused quite narrowly on spectral analysis. It is very useful for our research group. But we haven't put much effort into making it widely accessible or comprehensive.

I agree that in principle it would be good to merge efforts. However, we say this often about this type of project, but in practice I'm not sure we have the time to really accomplish it. It would take significant developer effort. xr-scipy seems the most mature and well designed of the three packages, so that should probably be the one we focus efforts on if we go that route.

If it were not for caching the grid setup, then this would make sense as a built-in xarray method interpolate_at (to complement interpolate_na).

This functionality sounds similar to reindex. Is it confusing to add this to reindex with an additional method such as method='linear'?

@fujiisoup also note @rabernat's xrft package, which has some overlap with your xr-scipy package.

This looks nice. I personally think that it would be nicer if we merge these similar packages and make them more complete. But yes, we can start listing these packages on a doc page first.

If it were not for caching the grid setup, then this would make sense as a built-in xarray method interpolate_at (to complement interpolate_na).

With caching, this feels a little bit beyond the standard xarray data model, but well suited for a companion package. It would be great to start listing these on a doc page, so users can easily find them!

@fujiisoup also note @rabernat's xrft package, which has some overlap with your xr-scipy package.

I think Interpolation (and scipy integration) is one of a good extensions of xarray ecosystem.

We discussed previously about contribute package/module in #1850 #1288 . It sounds a good idea to me to start from contrib.scipy module.

I notice that different people have developed almost the same package in parallel, https://github.com/lamorton/SciPyXarray https://github.com/fujiisoup/xr-scipy and maybe more.