Slow performance of pandas timestamp vs datetime

Question

I seem to be running into unexpectedly slow performance of arithmetic operations on pandas.Timestamp vs python regular datetime() objects.
Here is a benchmark that demonstrates:

import datetime
import pandas
import numpy

# using datetime:
def test1():
    d1 = datetime.datetime(2015, 3, 20, 10, 0, 0)
    d2 = datetime.datetime(2015, 3, 20, 10, 0, 15)
    delta = datetime.timedelta(minutes=30)

    count = 0
    for i in range(500000):
        if d2 - d1 > delta:
            count += 1

# using pandas:
def test2():
    d1 = pandas.datetime(2015, 3, 20, 10, 0, 0)
    d2 = pandas.datetime(2015, 3, 20, 10, 0, 15)
    delta = pandas.Timedelta(minutes=30)

    count = 0
    for i in range(500000):
        if d2 - d1 > delta:
            count += 1

# using numpy
def test3():
    d1 = numpy.datetime64('2015-03-20 10:00:00')
    d2 = numpy.datetime64('2015-03-20 10:00:15')
    delta = numpy.timedelta64(30, 'm')

    count = 0
    for i in range(500000):
        if d2 - d1 > delta:
            count += 1


  time1 = datetime.datetime.now()
  test1()
  time2 = datetime.datetime.now()
  test2()
  time3 = datetime.datetime.now()
  test3()
  time4 = datetime.datetime.now()

  print('DELTA test1: ' + str(time2-time1))
  print('DELTA test2: ' + str(time3-time2))
  print('DELTA test3: ' + str(time4-time3))

And corresponding results on my machine (python3.3, pandas 0.15.2):

DELTA test1: 0:00:00.131698
DELTA test2: 0:00:10.034970
DELTA test3: 0:00:05.233389

Is this expected?
Are there ways to eliminate the performance problem other than switching code to Python's default datetime implementation as much as possible?

Answer 1

I've got similar results on my machine:

$ python -mtimeit -s "from datetime import datetime, timedelta; d1, d2 = datetime(2015, 3, 20, 10, 0, 0), datetime(2015, 3, 20, 10, 0, 15); delta = timedelta(minutes=30)" "(d2 - d1) > delta"
10000000 loops, best of 3: 0.107 usec per loop
$ python -mtimeit -s "from numpy import datetime64, timedelta64; d1, d2 = datetime64('2015-03-20T10:00:00Z'), datetime64('2015-03-20T10:00:15Z'); delta = timedelta64(30, 'm')" "(d2 - d1) > delta"
100000 loops, best of 3: 5.35 usec per loop
$ python -mtimeit -s "from pandas import Timestamp, Timedelta; d1, d2 = Timestamp('2015-03-20T10:00:00Z'), Timestamp('2015-03-20T10:00:15Z'); delta = Timedelta(minutes=30)" "(d2 - d1) > delta"
10000 loops, best of 3: 19.9 usec per loop

datetime is several times faster than corresponding numpy, pandas analogs.

$ python -c "import numpy, pandas; print(numpy.__version__, pandas.__version__)"
('1.9.2', '0.15.2')

It is not clear why the difference is so large. It is true that numpy, pandas code is optimized for vectorized operations. But it is not obvious why these particular scalar operations are two orders of magnitude slower e.g., the adding of the explicit timezone does not slowdown datetime.datetime code:

$ python3 -mtimeit -s "from datetime import datetime, timedelta, timezone; d1, d2 = datetime(2015, 3, 20, 10, 0, 0, tzinfo=timezone.utc), datetime(2015, 3, 20, 10, 0, 15, tzinfo=timezone.utc); delta = timedelta(minutes=30)" "(d2 - d1) > delta"
10000000 loops, best of 3: 0.0939 usec per loop

To workaround the issue, you could try to convert native date/time types en masse into simpler (faster) analogs (e.g., POSIX timestamp represented as a float) if you can't use vectorized operations.