bug#8357: coreutils bug in date

[Bob Proulx]

tags 8357 + moreinfo notabug
thanks

Heikki Orsila wrote:
> Hello, I found a case where GNU date works pretty oddly.
> At 2011-03-27 Finland moved to 'summer time'. After midnight at 2011-03-28
> our web site's cron script did (date -I -d yesterday) which yielded
> unexpected results:
> 
> $ date
> Mon Mar 28 00:36:07 EEST 2011
> $ date -I -d yesterday
> 2011-03-26
> $ date -I -d '24 hours ago'
> 2011-03-26
> 
> Our scripts produced bad results due to this "feature/bug". Maybe our fault, 
> but this is definitely unexpected behaviour.

Thank you for your bug report.  However I this is not a bug in date
but a misunderstanding of how relative times and DST interact.

Note that 'date -I' is a deprecated option.  Please consider updating
the syntax.  In a future release -I may be removed.

First, let's understand when the time changes.

  $ zdump -v Europe/Helsinki | grep 2011
  Europe/Helsinki  Sun Mar 27 00:59:59 2011 UTC = Sun Mar 27 02:59:59 2011 EET 
isdst=0 gmtoff=7200
  Europe/Helsinki  Sun Mar 27 01:00:00 2011 UTC = Sun Mar 27 04:00:00 2011 EEST 
isdst=1 gmtoff=10800
  Europe/Helsinki  Sun Oct 30 00:59:59 2011 UTC = Sun Oct 30 03:59:59 2011 EEST 
isdst=1 gmtoff=10800
  Europe/Helsinki  Sun Oct 30 01:00:00 2011 UTC = Sun Oct 30 03:00:00 2011 EET 
isdst=0 gmtoff=7200

Sun Mar 27 02:59:59 2011 EET was the last second in EET and Sun Mar 27
04:00:00 2011 EEST was the very next second in your timezone.  There
is no 03:00:00 through 03:59:59.  The time skips (by law and politics)
directly to 04:00:00 in that timezone.  The day is defined to be one
hour shorter that day.

The documentation for date says:

     The unit of time displacement may be selected by the string `year'
  or `month' for moving by whole years or months.  These are fuzzy units,
  as years and months are not all of equal duration.  More precise units
  are `fortnight' which is worth 14 days, `week' worth 7 days, `day'
  worth 24 hours, `hour' worth 60 minutes, `minute' or `min' worth 60
  seconds, and `second' or `sec' worth one second.  An `s' suffix on
  these units is accepted and ignored.

     The string `tomorrow' is worth one day in the future (equivalent to
  `day'), the string `yesterday' is worth one day in the past (equivalent
  to `day ago').

So 'yesterday' or '24 hours ago' are equivalent.  These are exactly 24
hours ago or 86400 seconds.  This is a critical point in your problem.

  $ TZ=Europe/Helsinki date -d 'Mon Mar 28 00:36:07 EEST 2011'
  Mon Mar 28 00:36:07 EEST 2011

  $ TZ=Europe/Helsinki date -d 'Mon Mar 28 00:36:07 EEST 2011 24 hours ago'
  Sat Mar 26 23:36:07 EET 2011

This recreates the same results you posted.  The reasoning is a little
more obvious if I list out the hours explicitly.

  Mon Mar 28 00:36:07 EEST 2011   -0 hours
  Sun Mar 27 23:36:07 EEST 2011   -1 hours
  Sun Mar 27 22:36:07 EEST 2011   -2 hours
  Sun Mar 27 21:36:07 EEST 2011   -3 hours
  Sun Mar 27 20:36:07 EEST 2011   -4 hours
  Sun Mar 27 19:36:07 EEST 2011   -5 hours
  Sun Mar 27 18:36:07 EEST 2011   -6 hours
  Sun Mar 27 17:36:07 EEST 2011   -7 hours
  Sun Mar 27 16:36:07 EEST 2011   -8 hours
  Sun Mar 27 15:36:07 EEST 2011   -9 hours
  Sun Mar 27 14:36:07 EEST 2011  -10 hours
  Sun Mar 27 13:36:07 EEST 2011  -11 hours
  Sun Mar 27 12:36:07 EEST 2011  -12 hours
  Sun Mar 27 11:36:07 EEST 2011  -13 hours
  Sun Mar 27 10:36:07 EEST 2011  -14 hours
  Sun Mar 27 09:36:07 EEST 2011  -15 hours
  Sun Mar 27 08:36:07 EEST 2011  -16 hours
  Sun Mar 27 07:36:07 EEST 2011  -17 hours
  Sun Mar 27 06:36:07 EEST 2011  -18 hours
  Sun Mar 27 05:36:07 EEST 2011  -19 hours
  Sun Mar 27 04:36:07 EEST 2011  -20 hours  <-- just after change
  Sun Mar 27 02:36:07 EET 2011   -21 hours  <-- just before change
  Sun Mar 27 01:36:07 EET 2011   -22 hours
  Sun Mar 27 00:36:07 EET 2011   -23 hours
  Sat Mar 26 23:36:07 EET 2011   -24 hours

Observe that there is no 03:00:00-03:59:59 and March 27 is only 23
hours long that day.  It is shorter than 24 hours.  Using a 24 hour
calculation near midnight skips over that day completely.

With this table it should be obvious that date is behaving correctly
and the day that is shorter than 24 hours due to the conversion to DST
has caused -24 hours to skip completely over it.

Please see this FAQ entry for detailed explanation and hints on how to
avoid this problem.

  
http://www.gnu.org/software/coreutils/faq/#The-date-command-is-not-working-right_002e

Instead of doing calculations around DST I suggest using calculations
around noon.  I am not aware of any location that changes DST at noon
which makes it a much safer time to do date calculations.  The FAQ
entry provides examples.  Or if possible use UTC which also avoids the
problem by never switching to DST.

Bob