Imagine some data like this:

python sensor_data = xr.DataArray(np.arange(6).reshape((3, 2)), coords=[ ('time', [0, 2, 3]), ('sensor', ['A', 'C']), ])

Let's say we now want to sample these sensors at some arbitrary points. We can use vectorized indexing do this: python sensor_data.sel({ 'sensor': xr.DataArray(['A', 'A', 'A', 'B', 'C'], dims=['sample']), 'time': xr.DataArray([0, 1, 2, 0, 0], dims=['sample']) })

This fails because we are sampling one of our sensors at time 1, where we don't have any observations. We can add method='ffill' to fix this: ```python sensor_data.sel({ 'sensor': xr.DataArray(['A', 'A', 'A', 'B', 'C'], dims=['sample']), 'time': xr.DataArray([0, 1, 2, 0, 0], dims=['sample']) }, method='ffill')

array([0, 0, 2, 0, 1])

```

The problem is that the bogus sensor "B" is now getting a value ffilled from sensor "A"'s time 0 observation, which doesn't make a lot of sense because sensor names are arbitary. What we really want to do is apply the ffill only down the "time" array, so the sel call sitll fails if a sensor name is unknown but we can still benefit from ffilling in places where it makes sense.

So, it would be nice if we could supply a per-dimension method (or tolerance) like this:

python sensor_data.sel({ 'sensor': xr.DataArray(['A', 'A', 'A', 'B', 'C'], dims=['sample']), 'time': xr.DataArray([0, 1, 2, 0, 0], dims=['sample']) }, method={'time': 'ffill'})

From an implementation point of view, this looks like an easy addition in indexing.remap_label_indexers: https://github.com/pydata/xarray/blob/235b2e5bcec253ca6a85762323121d28c3b06038/xarray/core/indexing.py#L243