Hi, thanks for your report. Can you provide a minimal reproducible example? See https://matthewrocklin.com/minimal-bug-reports for an explanation

Hi, of course. The code above is as minimal as i could get it though. Oh wait, upon rechecking I could reduce the input data little bit further, sorry! Now, bit more reduced and without the comments.

import pandas as pd

data = pd.DataFrame([["C","B","B"],["B","A","A"],["B","A","B"]], columns=["0","1","2"])

data["0"] = data["0"].astype("category")
data["0"] = data["0"].cat.rename_categories({"C":"B", "B":"C"})

a = data.groupby(by=["0","1"])["2"].value_counts()
b = data.groupby(by=["0","1"]).size()
a.div(b)

For the expected behavior, but with a differently ordered index, use:

a.div(b.sort_index(ascending=False))

My findings so far:

• Removing any further column / row / or category-within-a-column of the DataFrame 'fixes' the unexpected behavior.
• One column must be of type category as the categories must be renamed in such way that alphabetical order of the categories itself are flipped.
• The column order given to the groupby(by) and [index] matters. I can only reproduce the unexpected behavior with order 0,1,2 & 1,0,2.
• Replacing a = data.groupby(by=["0","1"])["2"].value_counts() with a = data.groupby(by=["0","1","2"]).size() cause the math operation to behave as expected.

Based on the above, I would guess that rename_categories() causes value_counts() to return a 'corrupted' index.
I got stuck trying to investigate further by looking up the MultiIndex code in the repository, but I havent been able to locate it with a quick search.

Ah it's not only the data but also the necessary steps. No need to include groupby, you could simply create the DataFrames that groupby returns directly without any additional methods.

Edit: To be more clear, you can just write the MultiIndex down the way it's returned by the groupby op.

I actually tried! But those 'manually created' indexes behave as expected.

I did have a quick look what could be the difference between the 'manual' index and the one generated using the code snippet above. If i remember correctly the MultiIndex.equals() returned False, but the MultiIndex._values were perfectly identical.

Ah, this is a valuable information for your initial post. This points to something in groupby not working as expected.

Your example works as expected on main btw. I recall that we fixed something similar a couple of months ago (the groupby case).

Thanks for checking. (I tried to build main, but issue #51499 prevented me.)
In that case feel free the close this, This was a fun puzzle 😄

This might need tests, not sure

take