What is your issue?

Currently, the 'limit' argument of interpolate_na shows some counterintuitive/undocumented behaviour. Take the following example: python import xarray as xr import numpy as np n=np.nan da=xr.DataArray([n, n, n, 4, 5, n ,n ,n], dims=["y"]) da.interpolate_na('y', limit=1, fill_value='extrapolate') This will produce the following result: array([ 1., nan, nan, 4., 5., 6., nan, nan]) Two things are surprising, in my opinion:

1. The interpolated value 1 at the beginning is far from any of the given values
2. The filling is done only towards the 'right'. This asymmetric behaviour is not mentioned in the documentation.

Comparison to pandas

Similar behaviour can be created using pandas with the following arguments: python da=xr.DataArray([n, n, n, 4, 5, n ,n ,n], dims=["y"]) dap=da.to_pandas() dap.interpolate(method='slinear', limit=1, limit_direction='forward', fill_value='extrapolate')

This is equivalent to the current xarray behaviour, except there is no 1 at the beginning.

Cause

Currently, the fill mask in xarray is implemented using a rolling window operation, where values outside the array are assumed to be valid (therefore the 1). See xarray.core.missing._get_valid_fill_mask

Possible Solutions

Boundary Issue

Concerning the 1 at the beginning: I think this should be considered a bug. It is likely not what you would expect if you specify a limit. As stated, pandas does not create it as well.

Asymmetric Filling

Concerning the asymmetric filling, I see two options: 1. No changes to the code, but mention in the documentation that (effectively), a forward-fill is done. 2. Make something similar to what pandas is doing. In pandas, there are two additional arguments controlling the limit behaviour: limit_direction is controlling the fill direction (left, right or both). limit_area effectively controls if we only do interpolation or allow for extrapolation as well.

What do you think?

What is your issue?

If an array contains np.inf and a rolling operation is applied, all values after this one are nan if numbagg is used. Take the following example:

python import xarray as xr import numpy as np xr.set_options(use_numbagg=False) da=xr.DataArray([1,2,3,np.inf,4,5,6,7,8,9,10], dims=['x']) da.rolling(x=2).sum() Output <xarray.DataArray (x: 11)> Size: 88B array([nan, 3., 5., inf, inf, 9., 11., 13., 15., 17., 19.]) Dimensions without coordinates: x With Numbagg: python xr.set_options(use_numbagg=True) da=xr.DataArray([1,2,3,np.inf,4,5,6,7,8,9,10], dims=['x']) print(da.rolling(x=2).sum()) Output <xarray.DataArray (x: 11)> Size: 88B array([nan, 3., 5., inf, inf, nan, nan, nan, nan, nan, nan]) Dimensions without coordinates: x

What did I expect?

I expected no user-visible changes in the output values if numbagg is activated.

Maybe, this is not a bug, but expected behaviour for numbagg. The following warning was raised from the second call: .../Local/virtual_environments/xarray_performance/lib/python3.10/site-packages/numbagg/decorators.py:247: RuntimeWarning: invalid value encountered in move_sum return gufunc(*arr, window, min_count, axis=axis, **kwargs)

If this is expected, I think it would be good to have a page in the documentation which lists the downsides and limitations of the various tool to accelerate xarray. From the current installation docs, I assumed I just need to install numbagg/bottleneck to make xarray faster without any changes in output values.

Environment

xarray==2024.2.0 numbagg==0.8.0

• [x] Closes #7665
• [x] Tests added
• [ ] User visible changes (including notable bug fixes) are documented in whats-new.rst

This is an attempt to close #7665 and combine the current possibilities from xarray (max_gap) and pandas (limit_direction, limit_area) regarding interpolation of nan values. Please see also my comments in #7665 for the motivation. This PR already involves a full implementation, documentation and corresponding tests, but before any final polishing, I want to hear your thoughts. Specifically, I think the API and default options need to be discussed. (See the proposed documentation of DataArray.interpolate_na() / Dataset.interpolate_na() for the current state)

Implementation: Basically, I use ffill and bfill to calculate the coordinate of the left/right edge for every gap in the data. Based on edge coordinates, all masks (limit, limit_area, max_gap) are created.

On the long term, it might be interesting to provide those arguments to other na-filling methods as well (ffill, bfill, fillna).

Things to consider

limit_direction=forward

Pros: - Backward compatible: If limit is not None, this is the current behaviour (see #7665) - Pandas compatible: Forward is the pandas default.

Cons: - limit_direction=both feels more natural as default. If the user does interpolate_na('x', fill_value='extrapolate'), in my opinion they will expect all nans to be filled, including both boundaries. In contrast to pandas, this was the case in xarray before, but not anymore now if we follow pandas and set limit_direction=forward. both would also increase performance, since no restrictions need to be applied.

limit_use_coordinates=False

Pros: - Backward compatible - Pandas compatible -> Both xarray and pandas have no support for coordinate based limits so far.

Cons: - Inconsistent with the current default of use_coordinates=True

Generally, one might discuss if this separate argument is necessary or only one argument use_coordinates is sufficient. Imo, if the grid is irregular and use_coordinates=True, there is not a lot of sense in specifying the limit as a fixed number of grid cells. Alternatively, we could allow a three-tuple like use_coordinates=(True, True, False) to specify the index for interpolation, limit and max_gap separately (or something similar).

use_coordinates=True

So far, if there is no coordinate for dim, interpolation will succeed, falling silently back to a linearly increasing index. I feel, for use_coordinate=True, we should fail and inform the user to set use_coordinate=False if they really want a linear index. However, this is a breaking change. Maybe we can keep this behaviour with use_coordinate=None as new default option (= True if coord existent, else linear).

Performance

On my machine, the new limit implementation based on ffill/bfill seems to be a little less performant (10%) than the old one (based on rolling). There might be potential for improvements.