pd.concat merges object-datetime and np.datetime columns incorrectly.

[dalejung]

rows = []
rows.append([pd.datetime(2010, 1, 1), 1])
rows.append([pd.datetime(2010, 1, 2), 'hi']) # test col upconverts to obj

df2_obj = pd.DataFrame.from_records(rows, columns=['date', 'test'])

ind = pd.date_range(start="2000/1/1", freq="D", periods=10)
df1 = pd.DataFrame({'date': ind, 'test':range(10)})

cdf = pd.concat([df1, df2_obj])
cdf

The rows where date was a np.datetime are converted to ints, while the object dates are left as datetimes.

I'm not sure if this is a bug. I would expect the np.datetime column to upconvert into Timestamp objects?

I ran into this from a test failure. Thing is, it's a fairly old test and it used to pass. I checked it against 0.9.1 and 0.10.0 which failed as well. So I'm not sure if I was on a previous commit that worked differently. it's weird.

Either way, just wanted to get a check on what the expected behavior is.

EDIT: http://nbviewer.ipython.org/4428511/

[dalejung]

Just a note. This isn't an exact replica of my unit test. In my code, df1.date is created with a list of Timestamps. Which is why it was working before. Prior to e06334a
all of the date columns would've remained dtype(object).

[wesm]

fixed, thanks