BUG: setitem with at not inferring dtype correctly

[jreback]

xref #14179
xref #12499
xref #18027

http://stackoverflow.com/questions/23250333/in-pandas-dataframe-after-i-set-an-entire-column-i-cant-update-another-column

but .loc works

In [50]: df = pd.DataFrame(index=pd.date_range(start,periods=1), columns=['timenow','Live'])

In [51]: df.at[start,'timenow'] = datetime.today() # initial value

In [52]: df
Out[52]: 
                              timenow Live
2014-04-01 2014-04-23 12:42:38.883082  NaN

[1 rows x 2 columns]

In [53]: df._data
Out[53]: 
BlockManager
Items: Index([u'timenow', u'Live'], dtype='object')
Axis 1: <class 'pandas.tseries.index.DatetimeIndex'>
[2014-04-01]
Length: 1, Freq: D, Timezone: None
ObjectBlock: [timenow, Live], 2 x 1, dtype: object

In [54]: df = pd.DataFrame(index=pd.date_range(start,periods=1), columns=['timenow','Live'])

In [55]: df.loc[start,'timenow'] = datetime.today() # initial value

In [56]: df
Out[56]: 
                              timenow Live
2014-04-01 2014-04-23 12:43:00.554915  NaT

[1 rows x 2 columns]

In [57]: df._data
Out[57]: 
BlockManager
Items: Index([u'timenow', u'Live'], dtype='object')
Axis 1: <class 'pandas.tseries.index.DatetimeIndex'>
[2014-04-01]
Length: 1, Freq: D, Timezone: None
DatetimeBlock: [timenow, Live], 2 x 1, dtype: datetime64[ns]

[heaven00]

Hey, I am gonna give it a shot 👍

EDIT1: The behaviour on the StackOverflow link isn't replicated with the current build

In [1]: import pandas as pd

In [2]: from datetime import datetime

In [3]: start = pd.to_datetime('20140401')

In [4]: df = pd.DataFrame(index=pd.date_range(start,periods=1), columns=['timenow','Live'])

In [5]: df.at[start,'timenow'] = datetime.today() # initial value

In [6]: df
Out[6]: 
                               timenow Live
2014-04-01  2018-10-24 16:39:27.734368  NaN

In [7]: df.at[start,'timenow'] = datetime.today() # this works

In [8]: df
Out[8]: 
                               timenow Live
2014-04-01  2018-10-24 16:39:36.485698  NaN

In [9]: df.Live = True

In [10]: df.at[start,'timenow'] = datetime.today() # this doesn't work - nothing changes

In [11]: df
Out[11]: 
                               timenow  Live
2014-04-01  2018-10-24 16:39:55.149644  True

Also can you elaborate on the dtype not infered correctly?

[mroeschke]

This looks to be working on master. Could use a test.

[anirudnits]

@mroeschke I am working on this.
I have added the test test_datetimeField_after_setitem in pandas/tests/generic/test_frame.py and basically recorded the times after each operation(as mentioned in the above snippet) and asserted that neither of them are equal.
Seems good?

[mroeschke]

Sure, will be easier to review in a PR

[anirudnits]

@mroeschke sure I'll submit the PR soon.

[anirudnits]

@mroeschke submitted the PR, have a look.