I need to process a huge amount of CSV files where the time stamp is always a string representing the unix timestamp in milliseconds. I could not find a method yet to modify these columns efficiently.

This is what I came up with, however this of course duplicates only the column and I have to somehow put it back to the original dataset. I'm sure it can be done when creating the DataFrame?

import sys
if sys.version_info[0] < 3:
    from StringIO import StringIO
else:
    from io import StringIO
import pandas as pd

data = 'RUN,UNIXTIME,VALUE\n1,1447160702320,10\n2,1447160702364,20\n3,1447160722364,42'

df = pd.read_csv(StringIO(data))

convert = lambda x: datetime.datetime.fromtimestamp(x / 1e3)
converted_df = df['UNIXTIME'].apply(convert)

This will pick the column 'UNIXTIME' and change it from

0    1447160702320
1    1447160702364
2    1447160722364
Name: UNIXTIME, dtype: int64

into this

0   2015-11-10 14:05:02.320
1   2015-11-10 14:05:02.364
2   2015-11-10 14:05:22.364
Name: UNIXTIME, dtype: datetime64[ns]

However, I would like to use something like pd.apply() to get the whole dataset returned with the converted column or as I already wrote, simply create datetimes when generating the DataFrame from CSV.

You can do this as a post processing step using to_datetime and passing arg unit='ms':

In [5]:
df['UNIXTIME'] = pd.to_datetime(df['UNIXTIME'], unit='ms')
df

Out[5]:
   RUN                UNIXTIME  VALUE
0    1 2015-11-10 13:05:02.320     10
1    2 2015-11-10 13:05:02.364     20
2    3 2015-11-10 13:05:22.364     42