Unstacking produces flipped array when using decreasing coordinate values

[shoyer]

As reported by @robintw on the mailing list:

If we create a DataArray with the y co-ordinates decreasing:

da = xr.DataArray(np.random.rand(3, 4), dims = ('y', 'x'),
                  coords={'x': np.arange(4),
                          'y': np.arange(3, 0, -1)})

and then stack it and unstack it again:

stacked = da.stack(allpoints=['y', 'x'])
result = stacked.unstack('allpoints')

We find that the top-left value of the resulting array is not the same as the top-left value of the original array:

da.isel(x=0, y=0) # 0.1143
result.isel(x=0, y=0) # 0.696

However, the value at the 'bottom left' of the array matches the top left of the original array:

result.isel(x=0, y=2) # 0.1143

Replacing the y co-ordinates of da with increasing values gives the correct results:

da.coords['y'] = np.arange(da.coords.y)
da.isel(x=0, y=0) # 0.1143
result.isel(x=0, y=0) # 0.1143

I'm assuming this is unintentional behaviour - at least, I couldn't find anything in the docs about it. Is this meant to happen?

[shoyer]

The issue here is that pandas.MultiIndex.from_product (used in xarray's stack) sorts levels, which means that when we unstack we get back sorted levels. We should probably fix this by using a custom version of from_product that doesn't sort levels.