Allow dataset interpolation with different datatypes

[Illviljan]

Allow different datatypes (in particular booleans) to be interpolated instead of dropping them.

• Requires Faster interp #4740.
• Closes Dataset.interp drops boolean variables #4761
• Tests added
• Passes pre-commit run --all-files

[pep8speaks]

Hello @Illviljan! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2021-05-13 08:27:50 UTC

[dcherian]

This seems like a good change but shouldn't we do this only when method="nearest"?

[Illviljan]

How do you linearly interpolate strings, objects or booleans?

The problem for me is that we drop all the non-valid variables when I want linear interpolation for most of my stuff. One can argue it could be either of bfill, ffill or nearest but I chose nearest because that's an option for both missing.interp and reindex_variables. I suppose we can add it as an option? But what should the option be called when we want a backup interpolation method that does not modify the elementwise values? And should it then only be reindex_variables specific?

Another case I've been pondering how to handle is when numbers are used as IDs. Then it doesn't make sense to do interpolation that messes with the elements as well. But these integers are not easily determined though because you can use integers just for performance reasons and then wouldn't mind it becoming floats later. Maybe they will continue staying a special case.

[dcherian]

How do you linearly interpolate strings, objects or booleans?

This is why we drop them now. my point is that doing "nearest" when the user has specified method="linear" is v. confusing. Perhaps we should have method="linear+nearest" (or something better) for what you propose: linear for numeric, nearest for everything else.

Q: Does the current version of xarray just ignore strings, objects & booleans when method="nearest"?

[Illviljan]

Well, as a user I was very confused why my boolean variables silently disappeared.
And I don't get why ds.interp has any mandate to drop variables, if it can't handle interpolating a variable with the method an error should've been raised.
But I'm a lazy user too so I of course would prefer to not have to deal with errors or figure out which variables were dropped and do multiple interpolation steps and finally merge the different datasets back together again. This is why I prefer ds.interp to do the only current method that makes sense for non-numerics rather than dropping them.

ds.interp currently only interpolates if var.dtype.kind in "uifc", so it ignores bools, strings, and objects for any method. Which made sense I think because the scipy interpolants always attempts to force to float. This is why I use reindex_variables as it both retains the dtype and handles strings/objects just fine.

I've added a parameter now method_for_non_numerics for these cases now, so now you can choose which reindex method you want.

[max-sixty]

Let's discuss this at the next dev meeting.

[dcherian]

Thanks @Illviljan Apologies for the delay.

I have some minor comments but this looks very nice!

[Illviljan]

@dcherian, renamed the parameter and fixed some errors due to the #5102 merge.

Docs failure seems unrelated.

[dcherian]

Thanks @Illviljan ! Great PR. Thank you for your patience

[dcherian]

Oops just noticed this is missing a whats-new note. Can you send in a PR with that note please?