Date and time handling seem to always be a tiny bit frustrating. This is not only because getting it right is a really hard thing to do but also because there are many different ways of looking at the problem. In python, there are lots of differen types and modules for handling dates and times, including the builtin datetime module, numpy's datetime64 system and the many different tools and abstractions that pandas provides on top.

import numpy as np
import pandas as pd
import datetime

This datetime.datetime objects is initialized based on seconds. For example, this will work:

dt64 = np.datetime64('2015-02-19T15:20:15')
datetime_cast = dt64.astype(datetime.datetime)
datetime_cast

datetime.datetime(2015, 2, 19, 14, 20, 15)

However, if you start using types and time bases other than second and minutes, the conversions will just not work, and return long objects.

dt64 = np.datetime64('2015-02-19T15:20:15','ns')  # nanoseconds
datetime_cast = dt64.astype(datetime.datetime)
datetime_cast

1424355615000000000L

I stumbled across this when working with pandas timeseries. pd.Datetimeindex objects use datetime64[ns] objects under the hood to that if for some reason you were to create a timerange using pandas and work with the underlying datetime64[ns] array. A conversion to datetime.datetime objects will no longer work directly.

Let's say we have this construct:

dates = pd.date_range('2012-1-1', periods=5, freq='D')
dates

<class 'pandas.tseries.index.DatetimeIndex'>
[2012-01-01, ..., 2012-01-05]
Length: 5, Freq: D, Timezone: None

dates_dt64 = dates.values
dates_dt64

array(['2012-01-01T01:00:00.000000000+0100',
       '2012-01-02T01:00:00.000000000+0100',
       '2012-01-03T01:00:00.000000000+0100',
       '2012-01-04T01:00:00.000000000+0100',
       '2012-01-05T01:00:00.000000000+0100'], dtype='datetime64[ns]')

and we want see represent a value as a datetime.datetime object. As long as we stay within pandas, we're fine:

ts = dates[0] # scalar timestamp
ts

Timestamp('2012-01-01 00:00:00', offset='D')

ts.to_datetime()

datetime.datetime(2012, 1, 1, 0, 0)

If we use the underlying values however:

dt64 = dates_dt64[0]  # scalar datetime64 object
dt = dt64.astype(datetime.datetime)  # try cast
dt

1325376000000000000L

We get stuck with a long because dates_dt64 base is in nanoseconds.

ts.dtype

'org_babel_python_eoe'

The fix is easy when you know this is going on:

dates_dt64_s = dates_dt64.astype('datetime64[s]')
dt64_s = dates_dt64_s[0]  # single datetime63 object
dt_seconds = dt64_s.astype(datetime.datetime)
dt_seconds

datetime.datetime(2012, 1, 1, 0, 0)

Thank god that for most use cases, pandas handles all of these weird casts and conversions beautifully so that you don't need to know any of the above in 99% of your uses.