Hm, the test case that makes this fail is:

assert_identical(
    sc.concat((a['x', 0], a['x', 1]), dim='x'),
    a['x', :2]
)

and that doesn't hold now because a['x', 0].coords['x'] is unaligned and then sc.concat((a['x', 0], a['x', 1]), dim='x').coords['x'] is unaligned, while a['x', :2].coords['x'] is aligned.

So unless we want to make a['x', 0].coords['x'] aligned, or remove the identity in the example above, the approach in this PR is not going to work.

xarray behaves the same as scipp if the unaligned coordinate is renamed to the same as the concatenation dimension:

import numpy as np
import xarray as xr

da1 = xr.DataArray(
    np.arange(4),
    {'x': np.arange(4), 'b': np.array('abc')},
    dims=('x',)
)
da2 = xr.DataArray(
    2 * np.arange(4),
    {'x': np.arange(4), 'b': np.array('abc')},
    dims=('x',)
)
da3 = xr.DataArray(
    np.arange(4),
    {'x': np.arange(4), 'b': np.array('123')},
    dims=('x',)
)
print(da1, da2 ,da3)

comb = xr.concat([da1, da2], dim='b')
print(comb)

print(comb + da3)

<xarray.DataArray (x: 4)>
array([0, 1, 2, 3])
Coordinates:
  * x        (x) int64 0 1 2 3
    b        <U3 'abc' <xarray.DataArray (x: 4)>
array([0, 2, 4, 6])
Coordinates:
  * x        (x) int64 0 1 2 3
    b        <U3 'abc' <xarray.DataArray (x: 4)>
array([0, 1, 2, 3])
Coordinates:
  * x        (x) int64 0 1 2 3
    b        <U3 '123'
<xarray.DataArray (b: 2, x: 4)>
array([[0, 1, 2, 3],
       [0, 2, 4, 6]])
Coordinates:
  * x        (x) int64 0 1 2 3
  * b        (b) <U3 'abc' 'abc'
<xarray.DataArray (b: 2, x: 4)>
array([[0, 2, 4, 6],
       [0, 3, 6, 9]])
Coordinates:
  * x        (x) int64 0 1 2 3
  * b        (b) <U3 'abc' 'abc'

in the sense that b from d3 is dropped in favor of b from concat((d1, d2), dim='b').

Hm, the test case that makes this fail is:

assert_identical(
    sc.concat((a['x', 0], a['x', 1]), dim='x'),
    a['x', :2]
)

and that doesn't hold now because a['x', 0].coords['x'] is unaligned and then sc.concat((a['x', 0], a['x', 1]), dim='x').coords['x'] is unaligned, while a['x', :2].coords['x'] is aligned.

So unless we want to make a['x', 0].coords['x'] aligned, or remove the identity in the example above, the approach in this PR is not going to work.

• a['x', 0].coords['x'] needs to be unaligned since otherwise many ops such as a / a['x', 0] will fail.
• concat was designed to make the coord aligned again in this case, and I think that makes sense for such "dimension coords" (Xarray terminology)?

concat was designed to make the coord aligned again in this case, and I think that makes sense for such "dimension coords" (Xarray terminology)?

Yes this is how I understand it too. Xarray behaves same as scipp when the coord is a 'dimension coord', and different otherwise. But scipp doesn't have the concept of a dimension coord', so we have to choose one behavior.

It seems to me we can't do much about this.

concat was designed to make the coord aligned again in this case, and I think that makes sense for such "dimension coords" (Xarray terminology)?

Yes this is how I understand it too. Xarray behaves same as scipp when the coord is a 'dimension coord', and different otherwise. But scipp doesn't have the concept of a dimension coord', so we have to choose one behavior.

Well, if the coord has the same name as the dimension then it is a dimension coord. Others are not.

It seems to me we can't do much about this.

Can't we not set the alignment flag for non-dimension coords (that where not aligned in the input)?

Can't we not set the alignment flag for non-dimension coords (that where not aligned in the input)?

Aha I didn't know we could take the name of the coord into account, in that case yeah we can do that. I'll try it out

Not sure if we want to make the coord aligned if all instances of the coord in the inputs are aligned, or if we want to make the coord aligned if any instances of the coord in the input are aligned.

When all instances of the coord in the input have the same alignment then all and any are going to produce the same result.

If the different instances of the coord in the input have different alignment, should we prefer to set the coord on the concatenated array aligned or unaligned?

Not sure if we want to make the coord aligned if all instances of the coord in the inputs are aligned, or if we want to make the coord aligned if any instances of the coord in the input are aligned.

If any, because we should never remove and "aligned" flag, since users may rely on it for safety.

When all instances of the coord in the input have the same alignment then all and any are going to produce the same result.

If the different instances of the coord in the input have different alignment, should we prefer to set the coord on the concatenated array aligned or unaligned?

I think this may have worked correctly before the alignment flag was introduced (when coords meant "aligned"andattrs` meant "unaligned). Please check if the behavior after this fix brings us back to how it worked before: https://github.com/scipp/scipp/pull/3153/files#diff-ab8d4f264378e9dc974da66797785e393da2f84544f571ba8b5a5394c1dffc45. I think it is correct now, as it was?

Edit: Have to correct myself, it looks like it was broken before, since meta was used in the code?