cube.aggregated_by produces wrong bounds in some cases

[pp-mo]

Try the following:

import iris
import iris.cube
import numpy as np

cube = iris.cube.Cube(np.zeros((12, 2, 2)), standard_name='wind_speed', units='m s-1')
co_y = iris.coords.AuxCoord([1996, 1997, 1997, 1996, 1997, 1997, 1996, 1997, 1997, 1996, 1997, 1997], long_name='year')
co_m = iris.coords.AuxCoord(['Dec', 'Jan', 'Feb']*4, long_name='month')
co_r = iris.coords.AuxCoord([2, 2, 2, 3, 3, 3, 4, 4, 4, 1, 1, 1], standard_name=u'realization')
cube.add_aux_coord(co_y, 0)
cube.add_aux_coord(co_m, 0)
cube.add_aux_coord(co_r, 0)

print(cube)

means = cube.aggregated_by(['month','year'],iris.analysis.MEAN)
print('means.coord(\'realization\'):\n', means.coord('realization'))

Result is:

>>> print cube
wind_speed / (m s-1)                (-- : 12; -- : 2; -- : 2)
     Auxiliary coordinates:
          month                         x        -       -
          realization                   x        -       -
          year                          x        -       -
>>> 
>>> means = cube.aggregated_by(['month','year'],iris.analysis.MEAN)
>>> print 'means.coord(\'realization\'):\n', means.coord('realization')
means.coord('realization'):
AuxCoord(array([ 1.5,  1.5,  1.5]), bounds=array([[2, 1],
       [2, 1],
       [2, 1]]), standard_name=u'realization', units=Unit('1'))
>>> 

But we expected

• all points == 2 (instead of 1.5)
• all bounds == [1, 4] (instead of [2, 1])

Note:
Trying subsets of the data suggests that it is the last 3 realization values being out-of-order which triggers the problem...

>>> means = cube[:9].aggregated_by(['month','year'],iris.analysis.MEAN)
>>> print means.coord('realization')
AuxCoord(array([ 3.,  3.,  3.]), bounds=array([[2, 4],
       [2, 4],
       [2, 4]]), standard_name=u'realization', units=Unit('1'))

>>> means = cube[6:].aggregated_by(['month','year'],iris.analysis.MEAN)
>>> print means.coord('realization')
AuxCoord(array([ 2.5,  2.5,  2.5]), bounds=array([[4, 1],
       [4, 1],
       [4, 1]]), standard_name=u'realization', units=Unit('1'))

[rcomer]

This has tripped me up a few times, and I had thought that the solution would be to make sure the bounds are always of the form [min, max]. So I started working on a branch. However, some of the existing tests for aggregated_by are expecting to find bounds in the "wrong" order:

• tests.unit.cube.test_Cube.py line 500
• this CML file, from line 6, which is used at tests.test_aggregate_by line 281

So, what is the desired behaviour?

[rcomer]

#4315 fixes the bounds problem. The OP also notes that the points are wrong, and I have opened #4455 to more specifically cover that problem. So I think this can be closed.