homepage

Some common python utilities had problems on Feb 29 this year when
parsing dates using format strings that did not include a year in them.
>>> time.strptime('Feb 29', '%b %d')
Traceback (most recent call last):
 File "<stdin>", line 1, in ?
 File "/usr/lib/python2.4/_strptime.py", line 425, in strptime
 julian = datetime_date(year, month, day).toordinal() - \
ValueError: day is out of range for month
This is apparently because python assumes the year is 1900 unless it
explicitly parses another year out of the string.
Applications can work around this by always adding a year and a %Y to
the string they are parsing.
But not all date manipulating applications care about years. In this
case the application was fail2ban, bug report and patches to it here:
 http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=468382
Should the year default to 1900 (the equivalent of what the much more
forgiving C API does by leaving struct tm tm_year = 0) or should this
error be raised? If the answer is yes, works as is this is easy and
just turns into us adding a note in the documentation to mention the
behavior.
I do believe this was a valid bug in fail2ban as assuming the current
year for date parsing is a bad idea and will do the wrong thing when
parsing across a year change.
Python's strptime is much more strict than C strptime (glibc's C
strptime is happy to return tm_mon 2 tm_mday 31. Its range checking is
minimal.
here's a C test case to play with its behavior:
#include <assert.h>
#include <stdio.h>
#include <time.h>
int main(int argc, char *argv[]) {
 unsigned long ret, parsed;
 assert(argc == 2);
 struct tm tm = { 0 };
 ret = strptime(argv[1], "%b %d", &tm);
 parsed = ret - (unsigned long)(argv[1]);
 printf("ret 0x%x parsed %d tm_mon %d tm_mday %d tm_year %d\n",
 ret, parsed,
 tm.tm_mon, tm.tm_mday, tm.tm_year);
}
% ./foo 'Feb 28'
ret 0xffffda8a parsed 6 tm_mon 1 tm_mday 28 tm_year 0
% ./foo 'Feb 29'
ret 0xffffda8a parsed 6 tm_mon 1 tm_mday 29 tm_year 0
% ./foo 'Feb 31'
ret 0xffffda8a parsed 6 tm_mon 1 tm_mday 31 tm_year 0
% ./foo 'Feb 32'
ret 0x0 parsed 9596 tm_mon 1 tm_mday 0 tm_year 0

The documentation already mentions that the default values when
information left out is (1900, 1, 1, 0, 0, 0, 0, 1, -1) so the docs are
already clear. If you want to generate a patch to make the default year
be this year I would be willing to review it and consider applying it. I
doubt very much code would break because of this.

Here is a patch, hope it'll make it to 2.6

Applying the _strptime.diff patch broke the _strptime
test("test_defaults"). Once you change the year, you also have to adapt
the day of week, as this becomes dynamic, too. The rest remains the
same, though. I attached a patch to this test which tests for the
new-years day of the current year instead of 1900, but I feel like
changing the semantic of the default value is no minor change. Also, I
am not sure what the documentation should say then.

After having thought about this I have decided I am going to stick with
the current semantics. Having the year change underneath code based
solely on when it executes will cause more problems than it will solve.