Custom ordering in a grouped bar chart not applied in Altair?

Question

I am trying to achieve a custom sort order for a series of categories in a grouped bar chart.

An issue on GitHub from a few years ago suggests how to do this: by supplying a list with the custom sort order to alt.Column. Supplying a list with a sort answer is also mentioned in Joel Ostblom's answer to this related Q.

Omitting the sort parameter from alt.Column() results in an alphabetical sorting:

Supplying a custom sort order results in no ordering / equivalent to sort=None I think, since the columns appear in the same order as they appear in the data:

sort=["easy to use", "easy to understand", "complete", "representative",
      "visually appealing", "b&l useful", "rating"]

Where am I going wrong in applying sort order?

My data is shaped as follows:

    timestamp           value   attribute   system
0   23/03/2024 16:28:33 4.0 representative  A
1   23/03/2024 16:28:33 5.0 representative  B
2   23/03/2024 16:28:33 4.0 representative  C
3   23/03/2024 16:28:33 3.0 easy to use A
4   23/03/2024 16:28:33 5.0 easy to use B
... ... ... ... ...
99  23/03/2024 19:24:41 5.0 b&l useful  B
100 23/03/2024 19:24:41 5.0 b&l useful  C
101 23/03/2024 19:24:41 3.0 rating  A
102 23/03/2024 19:24:41 5.0 rating  B
103 23/03/2024 19:24:41 5.0 rating  C

The attributes I am grouping by are:

d1['attribute'].unique()
array(['representative', 'easy to use', 'easy to understand', 'complete',
       'visually appealing', 'b&l useful', 'rating'], dtype=object)

and I create the bar plot via:

alt.Chart(d1).mark_bar().encode(
    alt.Y('mean(value):Q', scale=alt.Scale(domain=[1,5]),),
    alt.X('system:N', axis=alt.Axis(labelAngle=-0, title=None,),),
    alt.Column(
        'attribute:N',
        sort=["easy to use", "easy to understand", "complete", "representative",
              "visually appealing", "b&l useful", "rating"],
        header=alt.Header(title=None, labelOrient='bottom'),
        
    ),
    color='system:N',
    tooltip=['mean(value)', 'attribute', 'system'],
)

Answer 1

Workaround: use xOffset for a visually-similar result

Instead of using a Column grouping, it is possible to use xOffset (one of several channels available in Altair charts) to achieve a similar outcome.

alt.Chart(d1).mark_bar().encode(
     alt.Y('mean(value):Q', scale=alt.Scale(domain=[1,5]),),
     alt.X('attribute:N',
           axis=alt.Axis(labelAngle=-0, title=None,),
           sort=["easy to use", "easy to understand", "complete", "representative",
                "visually appealing", "b&l useful", "rating"],
          ),
     alt.XOffset('system:N'),
     color='system:N',
     tooltip=['mean(value)', 'attribute', 'system', 'count()'],
) 

The sort parameter is moved to the x-axis, where it produces the correct result- a custom sort order based on the supplied list.

It is also possible to reorder the data itself and let the sort be based on source data order.

Why a workaround?

As discussed on the related issue, there is a bug in Vega-Lite, which Atair is a python wrapper for. Quoting Joel Ostblom from there:

After some exploring, it seems to be an issue with aggregating and sorting facets at the same time. I had no idea this was an existing bug, but it has been reported in VL already here vega/vega-lite#5366; you might be able to ask for workaround there and thumbs up that issue.

The existing bug notes that Vega-Lite does not produce the correct ordering when supplied a custom order.

Answer 2

Doing the aggregation in a transform does work. Example using different data:

import altair as alt  
from vega_datasets import data

source = data.barley()

base = alt.Chart(source).mark_bar().encode(
    alt.Column('site:N',
               sort=["Waseca", "Morris", "University Farm", 
                     "Grand Rapids", "Crookston", "Duluth"]
              ),
    x='year:O',
    color='year:N',
)

chart1 = base.encode(y=alt.Y('sum(yield):Q'))
chart2 = base.transform_aggregate(y2="sum(yield)",groupby=['site', 'year']).encode(y=alt.Y('y2:Q'))

chart1 | chart2