Pandas weird behavior using .replace() to swap values

Question

I stumbled upon a weird and inconsistent behavior for Pandas replace function when using it to swap two values of a column. When using it to swap integers in a column we have

df = pd.DataFrame({'A': [0, 1]})
df.A.replace({0: 1, 1: 0})

This yields the result:

df
A
1
0

However, when using the same commands for string values

df = pd.DataFrame({'B': ['a', 'b']})
df.B.replace({'a': 'b', 'b': 'a'})

We get

df
B
'a'
'a'

Can anyone explain me this difference in behavior, or point me to a page in the docs that deals with inconsistencies when using integers and strings in pandas?

Answer 1

Yup, this is definitely a bug, so I've opened a new issue - GH20656.

It looks like pandas applies the replacements successively. It makes first replacement, causing "a" to be replaced with "b", and then the second, causing both "b"s to be replaced by "a".

In summary, what you see is equivalent to

df.B.replace('a', 'b').replace('b', 'a')

0    a
1    a
Name: B, dtype: object

Which is definitely not what should be happening.

---

There is a workaround using str.replace with a lambda callback.

m = {'a': 'b', 'b': 'a'}
df.B.str.replace('|'.join(m.keys()), lambda x: m[x.group()])

0    b
1    a
Name: B, dtype: object