sel(method='nan') option filling missing values with NaN

[frexvahi]

Rationale

I would find it helpful to be able to specify "give me NaNs when the selector is not in the corresponding coordinate rather than raising an IndexError". This could go as a method option e.g. method='nan'.

>>> data.dims
 ('time', 'feature')
>>> time_array.dims
 ('base_time', 'offset')
>>> data.sel(time=time_array, method='nan')
 <array with NaNs in place of missing times>

Currently I work around the problem by appending a row to the data array with time coordinate NaT and all values NaN, then replacing missing times from time_array with NaT. However, this gives a FutureWarning from xarray/core/duck_array_ops.py:138: In the future, 'NAT == x' and 'x == NAT' will always be False.

If this is something you would like to see then I could have a go at making a pull request for this.

[shoyer]

This is exactly what the .reindex() method is designed for. You can use it like sel, but it gives the appropriate NaN/NaT value instead of raising an error.

[frexvahi]

I tried .reindex() and got ValueError: IndexVariable objects must be 1-dimensional

so allowing reindex with a two-dimensional selector would be another solution.

[shoyer]

so allowing reindex with a two-dimensional selector would be another solution.

Yes, we want to support that. See #1553.

We're already written most of the underlying machinery, so it should be pretty easy to add now. There are some backwards compatibility concerns with adjusting this for 1D arrays prior to a breaking release (see this warning), but I don't see any reason why we can't enable this for higher dimensional indexers right now.

[frexvahi]

OK, that should solve my problem. I will go on using the NaT hack for now and refer back here as and when that stops working because NaT != NaT