Bars doesn’t respect dict ordering

[G-Guillard]

ALL software version info

Software Version Info

holoviews 1.21.0
bokeh 3.6.3, 3.8.0
python 3.11.11, 3.12.10

Description of expected behavior and the observed behavior

Since Python 3.7, dictionaries respect insertion ordering. As a consequence, when applying hv.Bars on a dict, one expects the bar plot to respect the dict ordering as well. However hv.Bars relies on the keys alphabetical or numerical ordering instead. Note that it behaves the same way with an explicit OrderedDict.

In addition, the documentation on how to use redim is somehow confusing.

Complete, minimal, self-contained example code that reproduces the issue

import holoviews as hv
from collections import Counter, OrderedDict

a = {'B': 9, 'A': 4, 'C': 1}
d = OrderedDict(a)

## don’t work -> ABC
hv.Bars(a)
# hv.Bars(d)

## work -> BAC
# hv.Bars(a, hv.Dimension("foobar")).redim.values(foobar=list(a.keys()))
# hv.Bars(a, hv.Dimension("foobar", values=list(a.keys())))
# hv.Bars(a).sort(by="y", reverse=True)
# hv.Bars(list(d.items()))

## some working alternatives
# b = sorted(a.items(), key=lambda x: x[1], reverse=True)
# hv.Bars(b)  # works -> BAC
# c = Counter(a).most_common()
# hv.Bars(c)  # works -> BAC

[hoxbro]

I don't think we can change this behavior, even though that is what you expect, it isn't necessary what others expect.

We could definitely improve the documentation about redim. Do you have any suggestions?

FWIW, the way I would use redim is likely like this: hv.Bars(a).redim.values(x=a)

[maximlt]

I don't think we can change this behavior, even though that is what you expect, it isn't necessary what others expect.

These days, respecting the input key order sounds like the appropriate and expected behavior, no?

[hoxbro]

These days, respecting the input key order sounds like the appropriate and expected behavior, no?

If we were adding this today, I would likely do it by input order. But we are not. People can have depended on this being sorted. I don't see an easy way to warn users about this behavior change without also being annoying.

I don't see either of them as the correct way; there are pros/cons for both.

[maximlt]

I think sorting a categorical dimension by name is incorrect. In this example with A/B/C it feels okay but in real-world examples, it's usually awkward. I can only assume the sorting was done because key insertion order wasn't preserved in dictionaries previously, but it is now. Maybe a 2.0 change?

[hoxbro]

I think sorting a categorical dimension by name is incorrect

Incorrect is such a strong stand - my opinion is there is no right answer. By sorting values, the order will also be deterministic if the dict is the output of some calculation based on a set, {pd, pl}.unique, or passing through two loops, which shuffle things around based on some third thing.

[maximlt]

the order will also be deterministic if the dict is the output of some calculation based on a set, {pd, pl}.unique, or passing through two loops, which shuffle things around based on some third thing.

Do you think of internal HoloViews use cases that could be affected by that? If it's about the user directly, well sure but that's true for many things and it's their responsibility. For example, pn.widgets.Tabulator(df) will display a table with the rows and columns ordered as in the dataframe. I'd certainly be not appropriate for Tabulator to present the columns alphabetically sorted.

[G-Guillard]

even though that is what you expect, it isn't necessary what others expect

Seems to me a bit far-fetched in this particular case, but okay. What about OrderedDict ? What proportion of users would expect carefully optimized OrderedDicts to be eventually alphabetically ordered anyway ?

We could definitely improve the documentation about redim. Do you have any suggestions?

I don’t really remember what I did wrong but it took some time to get a working line. Maybe just adding a few examples like mine and yours would help readers getting a grasp of how it works. Also maybe link to redim — although it wouldn’t have helped much here.

[philippjfr]

I do agree that in the long run it would be nice to get to a place where we maintain the order. How to get there is another matter though. I think we may simply have to do this in a major release, since I don't think there really is any way to warn users of this, so it would have to be called out very prominently in the release notes.