Adding a column to a DataFrame always creates a copy of a Series

I don’t know if this is a valid behaviour, but it seems to me like a bug?

>>> s = pd.Series([1,2,3])
>>> df = pd.DataFrame(s)
>>> df.index is s.index
True
>>> df.iloc[0, 0] = 33
>>> df
    0
0  33
1   2
2   3
>>> s
0    33
1     2
2     3
dtype: int64

So far so good.

But if I do:

>>> s = pd.Series([1,2,3])
>>> df = pd.DataFrame(index=s.index)
>>> df[0] = s
>>> df.index is s.index
True
>>> df.iloc[0, 0] = 33
>>> df
    0
0  33
1   2
2   3
>>> s
0    1
1    2
2    3
dtype: int64

Basically there is no way to add a column to the DataFrame without creating a copy of the data. This seems like a suboptimal behaviour since the operation:

df['c'] = df['a'] + df['b']

First create a Series object in the memory, and then create a copy of it that get’s assigned to the DataFrame column c.

I also understand why this can be a desired behaviour, so maybe this issue could be reformulated into a question: Is there a way to add a column to a DataFrame without creating a copy of the data.

[charlescbeebe]

+1 for deprecating copy from the public API – I also suggest that the special behavior of the assignment operator be prominently announced in the relevant docstrings and in 10 Minutes to pandas. You have to dig around quite a bit in the source to figure out that:

df['x'] = df['y']

is actually:

from copy import deepcopy
df['x'] = deepcopy(df['y'])

or (because of the redundancy in the public API):

df['x'] = df['y'].copy()

While I appreciate the argument that this case is special enough to break with the expected behavior of the language in which you’ve chosen to implement this library because the core devs perceive it as the default use case in this context, it is not such an obvious change that it’s reasonable to leave people to figure it out on their own.

[jreback]

its only possible to not copy in very limited circumstances (which IMHO are not necessary anyhow) so go ahead and close