Hello,

I was testing edge case behaviour of some code of mine and stumbled into this unexpected domain error from timedelta:

  >>> from datetime import *
  >>> timedelta(999999999, 86399, 999999) - timedelta(999999999, 86399, 999998)
  Traceback (most recent call last):
    File "<stdin>", line 1, in ?
  OverflowError: days=-1000000000; must have magnitude <= 999999999
  >>>

The expected result is of course timedelta(0, 0, 1):

  >>> timedelta(999999999, 86399, 999998) + timedelta(0, 0, 1)
  datetime.timedelta(999999999, 86399, 999999)

Both time deltas are within the range documented in
/usr/share/doc/python-doc/html/lib/datetime-timedelta.html

I could reproduce it on 2.6.6 and 3.1.3. I don't have access to other python versions.


Ciao,

Enrico


Note: I reported it 4 years ago in the Debian BTS (http://bugs.debian.org/408872) but I noticed now that the Debian maintainer doesn't seem to have bothered forwarding it here :(

To emphasize that it's only days parameter that is overflowing here is another case:

>>> timedelta(999999999, 1, 1) - timedelta(999999999, 0, 0)
datetime.timedelta(0, 1, 1)
>>> timedelta(999999999, 1, 1) - timedelta(999999999, 1, 0)
Traceback (most recent call last):
  File "<pyshell#18>", line 1, in <module>
    timedelta(999999999, 1, 1) - timedelta(999999999, 1, 0)
OverflowError: days=-1000000000; must have magnitude <= 999999999

What happens is the second value is negated (__neg__) which causes it to become less than timedelta.min and that is causing OverflowError.

> What happens is the second value is negated (__neg__)
> which causes it to become less than timedelta.min and
> that is causing OverflowError.

Yes, and running the test case without C acceleration makes this obvious:

>>> import sys
>>> sys.modules['_datetime'] = None
>>> from datetime import *
>>> timedelta(999999999, 86399, 999999) - timedelta(999999999, 86399, 999998)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "Lib/datetime.py", line 488, in __sub__
    return self + -other
  File "Lib/datetime.py", line 501, in __neg__
    -self._microseconds)
  File "Lib/datetime.py", line 426, in __new__
    raise OverflowError("timedelta # of days is too large: %d" % d)
OverflowError: timedelta # of days is too large: -1000000000

Attached patch fixes the issue.  I would like to think some more about C int overflow before committing.

New changeset 76180cc853b6 by Alexander Belopolsky in branch '3.2':
Issue #11576: Fixed timedelta subtraction glitch on big timedelta values
http://hg.python.org/cpython/rev/76180cc853b6

New changeset d492915cf76d by Alexander Belopolsky in branch 'default':
Issue #11576: Fixed timedelta subtraction glitch on big timedelta values
http://hg.python.org/cpython/rev/d492915cf76d

New changeset 202a9feb1fd6 by Alexander Belopolsky in branch '2.7':
Issue #11576: Fixed timedelta subtraction glitch on big timedelta values
http://hg.python.org/cpython/rev/202a9feb1fd6

I can still reproduce that error with Python 2.7.2:

Python 2.7.1 (r271:86832, Apr 27 2011, 15:42:42) 
[GCC 4.5.2] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> from datetime import *
>>> timedelta(999999999, 86399, 999999) - timedelta(999999999, 86399, 999998)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
OverflowError: days=-1000000000; must have magnitude <= 999999999
>>> 

and I see that patch is applied in 2.7.2.