Idiomatic way to parse POSIX timestamps in pandas?

Question

I have a csv file with a time column representing POSIX timestamps in milliseconds. When I read it in pandas, it correctly reads it as Int64 but I would like to convert it to a DatetimeIndex. Right now I first convert it to datetime object and then cast it to a DatetimeIndex.

In [20]: df.time.head()

Out[20]: 
0    1283346000062
1    1283346000062
2    1283346000062
3    1283346000062
4    1283346000300
Name: time

In [21]: map(datetime.fromtimestamp, df.time.head()/1000.)
Out[21]: 
[datetime.datetime(2010, 9, 1, 9, 0, 0, 62000),
 datetime.datetime(2010, 9, 1, 9, 0, 0, 62000),
 datetime.datetime(2010, 9, 1, 9, 0, 0, 62000),
 datetime.datetime(2010, 9, 1, 9, 0, 0, 62000),
 datetime.datetime(2010, 9, 1, 9, 0, 0, 300000)]

In [22]: pandas.DatetimeIndex(map(datetime.fromtimestamp, df.time.head()/1000.))
Out[22]: 
<class 'pandas.tseries.index.DatetimeIndex'>
[2010-09-01 09:00:00.062000, ..., 2010-09-01 09:00:00.300000]
Length: 5, Freq: None, Timezone: None

Is there an idiomatic way of doing this? And more importantly is this the recommended way of storing non-unique timestmaps in pandas?

Answer 1

You can use a converter in combination with read_csv.

In [423]: d = """\
timestamp data
1283346000062 a
1283346000062 b
1283346000062 c
1283346000062 d
1283346000300 e
"""

In [424]: fromtimestamp = lambda x:datetime.fromtimestamp(int(x) / 1000.)

In [425]: df = pandas.read_csv(StringIO(d), sep='\s+', converters={'timestamp': fromtimestamp}).set_index('timestamp')

In [426]: df.index
Out[426]:
<class 'pandas.tseries.index.DatetimeIndex'>
[2010-09-01 15:00:00.062000, ..., 2010-09-01 15:00:00.300000]
Length: 5, Freq: None, Timezone: None

In [427]: df
Out[427]:
                           data
timestamp
2010-09-01 15:00:00.062000    a
2010-09-01 15:00:00.062000    b
2010-09-01 15:00:00.062000    c
2010-09-01 15:00:00.062000    d
2010-09-01 15:00:00.300000    e

Answer 2

Internally, Timestamps are stored in int representing nanoseconds. They use the numpy datetime/timedelta. The issue with your timestamps is that they are in ms precision, which you already know since you're dividing by 1000. In this case, it's easier to astype('M8[ms]'). It's essentially saying view these ints as datetime-ints with ms precision.

In [21]: int_arr
Out[21]: 
array([1283346000062, 1283346000062, 1283346000062, 1283346000062,
       1283346000300])

In [22]: int_arr.astype('M8[ms]')
Out[22]: 
array(['2010-09-01T09:00:00.062-0400', '2010-09-01T09:00:00.062-0400',
       '2010-09-01T09:00:00.062-0400', '2010-09-01T09:00:00.062-0400',
       '2010-09-01T09:00:00.300-0400'], dtype='datetime64[ms]')

Pandas will assume any regular int array is in M8[ns]. An array with a datetime64 dtype will be correctly interpreted. You can view the M8[ns] representation of a DatetimeIndex by access ing it's asi8 attribute.

[EDIT] I realize that this won't help you directly with read_csv. Just thought I'd throw out how to quickly convert between timestamp arrays.