BUG #3431: age() gets the days wrong

"Pelle Johansson" <pelle@morth.org> writes:

The age() function seem to work by first counting months until less than a
month remains to to the second argument, then counting days left. This
doesn't give the correct result, as shown by this example:

age() isn't intended to give the same result as minus, so I'm not
convinced that this behavior is wrong. I'm more worried about what
applications we might break by changing it --- age() has always worked
the way it does now.

regards, tom lane

"Pelle Johansson" <pelle@morth.org> writes:

The age() function seem to work by first counting months until less than a
month remains to to the second argument, then counting days left. This
doesn't give the correct result, as shown by this example:

# select column1, age(column1, '2006-11-02'), date '2006-11-02' +
age(column1, '2006-11-02') from (values ('2007-01-31'::date),
('2007-02-01')) as alias;
column1 | age | ?column?
------------+----------------+---------------------
2007-01-31 | 2 mons 29 days | 2007-01-31 00:00:00
2007-02-01 | 2 mons 29 days | 2007-01-31 00:00:00
(2 rows)

I took another look at this example. I believe what is actually going
wrong here is that when timestamp_age converts a month into an
equivalent number of days, it uses the number of days in the first
month of the interval it's dealing with (ie, the month containing
the earlier of the two dates). This is just wrong, because interval
addition adds months first and then days. The appropriate conversion
to use is actually the length of the next-to-last month of the interval.

As an example, 8.2 and CVS HEAD produce

regression=# select age('2007-03-14', '2007-02-15');
age
---------
27 days
(1 row)

which is reasonable, but

regression=# select age('2007-04-14', '2007-02-15');
age
---------------
1 mon 27 days
(1 row)

is not so reasonable, nor is

regression=# select age('2007-03-14', '2007-01-15');
age
---------------
1 mon 30 days
(1 row)

If we change the code to use the next-to-last month of the interval
then these two cases produce '1 mon 30 days' and '1 mon 27 days'
respectively.

Another problem is that the code isn't doing the propagate-to-next-field
bit for negative fractional seconds. Hence it produces

regression=# select age('2007-02-14 01:00:00', '2007-01-15 01:00:00.4');
age
----------------------
30 days -00:00:00.40
(1 row)

which is maybe not incorrect, but certainly fairly inconsistent with

regression=# select age('2007-02-14 01:00:00', '2007-01-15 01:00:01');
age
------------------
29 days 23:59:59
(1 row)

Hence I propose the attached patch. This does not change any existing
regression test outputs, but it does change the example given in the
documentation: age(timestamp '2001-04-10', timestamp '1957-06-13')
will now produce '43 years 9 mons 28 days' not 27 days. Which actually
is correct if you try to add back the result to timestamp '1957-06-13'.
It also appears to fix Palle's example:

regression=# select column1, age(column1, '2006-11-02'), date '2006-11-02' +
age(column1, '2006-11-02') from (values ('2007-01-31'::date),
('2007-02-01')) as alias;
column1 | age | ?column?
------------+----------------+---------------------
2007-01-31 | 2 mons 29 days | 2007-01-31 00:00:00
2007-02-01 | 2 mons 30 days | 2007-02-01 00:00:00
(2 rows)

As I said earlier, I'm worried about changing the behavior of a function
that's been around for so long, so I'm disinclined to back-patch this.
But it seems like a reasonable change to make in 8.3. Comments?

regards, tom lane

Hi Tom,

I only have one remark really, which I first thought of after sending
the report.
If you were to use the result for subtracting from the first value,
instead of adding to the second, the conditions are reversed. It's
not really as obvious as I first thought whether there's 2 months and
29 days or 2 months and 30 days between 2006-11-02 and 2007-02-01...
If one want mathematical correctness, it will have to be defined
precisely.
--
Pelle Johansson

8 jul 2007 kl. 22.07 skrev Tom Lane:

Pelle Johansson <pelle@morth.org> writes:

If you were to use the result for subtracting from the first value,
instead of adding to the second, the conditions are reversed. It's
not really as obvious as I first thought whether there's 2 months and
29 days or 2 months and 30 days between 2006-11-02 and 2007-02-01...

Hmm, that's a really good point; perhaps the original author was
thinking of it in those terms, in which case using the first month of
the interval is indeed sane. (Almost: I believe that the loop can
iterate more than once, and then you need to look to the second month
etc. The code's not doing that, so there's still a corner-case bug,
plus the fsec issue.)

Other than that corner case, it seems the behavior we currently have is
if x > y, age() produces a positive interval such that
x - age(x, y) = y
if x < y, age() produces a negative interval such that
y + age(x, y) = x

Are we satisfied with just documenting that, or do we want to change it,
and if so to what?

As the code currently stands, we have the symmetry property
age(x,y) = - age(y,x)
for all x,y. I don't think we can preserve that if we try to simplify
the relationship to interval addition/subtraction.

Comments?

regards, tom lane

I don't see this as applied yet.

---------------------------------------------------------------------------

Tom Lane wrote:

"Pelle Johansson" <pelle@morth.org> writes:

The age() function seem to work by first counting months until less than a
month remains to to the second argument, then counting days left. This
doesn't give the correct result, as shown by this example:

# select column1, age(column1, '2006-11-02'), date '2006-11-02' +
age(column1, '2006-11-02') from (values ('2007-01-31'::date),
('2007-02-01')) as alias;
column1 | age | ?column?
------------+----------------+---------------------
2007-01-31 | 2 mons 29 days | 2007-01-31 00:00:00
2007-02-01 | 2 mons 29 days | 2007-01-31 00:00:00
(2 rows)

I took another look at this example. I believe what is actually going
wrong here is that when timestamp_age converts a month into an
equivalent number of days, it uses the number of days in the first
month of the interval it's dealing with (ie, the month containing
the earlier of the two dates). This is just wrong, because interval
addition adds months first and then days. The appropriate conversion
to use is actually the length of the next-to-last month of the interval.

As an example, 8.2 and CVS HEAD produce

regression=# select age('2007-03-14', '2007-02-15');
age
---------
27 days
(1 row)

which is reasonable, but

regression=# select age('2007-04-14', '2007-02-15');
age
---------------
1 mon 27 days
(1 row)

is not so reasonable, nor is

regression=# select age('2007-03-14', '2007-01-15');
age
---------------
1 mon 30 days
(1 row)

If we change the code to use the next-to-last month of the interval
then these two cases produce '1 mon 30 days' and '1 mon 27 days'
respectively.

Another problem is that the code isn't doing the propagate-to-next-field
bit for negative fractional seconds. Hence it produces

regression=# select age('2007-02-14 01:00:00', '2007-01-15 01:00:00.4');
age
----------------------
30 days -00:00:00.40
(1 row)

which is maybe not incorrect, but certainly fairly inconsistent with

regression=# select age('2007-02-14 01:00:00', '2007-01-15 01:00:01');
age
------------------
29 days 23:59:59
(1 row)

Hence I propose the attached patch. This does not change any existing
regression test outputs, but it does change the example given in the
documentation: age(timestamp '2001-04-10', timestamp '1957-06-13')
will now produce '43 years 9 mons 28 days' not 27 days. Which actually
is correct if you try to add back the result to timestamp '1957-06-13'.
It also appears to fix Palle's example:

regression=# select column1, age(column1, '2006-11-02'), date '2006-11-02' +
age(column1, '2006-11-02') from (values ('2007-01-31'::date),
('2007-02-01')) as alias;
column1 | age | ?column?
------------+----------------+---------------------
2007-01-31 | 2 mons 29 days | 2007-01-31 00:00:00
2007-02-01 | 2 mons 30 days | 2007-02-01 00:00:00
(2 rows)

As I said earlier, I'm worried about changing the behavior of a function
that's been around for so long, so I'm disinclined to back-patch this.
But it seems like a reasonable change to make in 8.3. Comments?

regards, tom lane

--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://www.enterprisedb.com

+ If your life is a hard drive, Christ can be your backup. +

Sorry, I see there was later discussion.

---------------------------------------------------------------------------

Tom Lane wrote:

"Pelle Johansson" <pelle@morth.org> writes:

The age() function seem to work by first counting months until less than a
month remains to to the second argument, then counting days left. This
doesn't give the correct result, as shown by this example:

# select column1, age(column1, '2006-11-02'), date '2006-11-02' +
age(column1, '2006-11-02') from (values ('2007-01-31'::date),
('2007-02-01')) as alias;
column1 | age | ?column?
------------+----------------+---------------------
2007-01-31 | 2 mons 29 days | 2007-01-31 00:00:00
2007-02-01 | 2 mons 29 days | 2007-01-31 00:00:00
(2 rows)

I took another look at this example. I believe what is actually going
wrong here is that when timestamp_age converts a month into an
equivalent number of days, it uses the number of days in the first
month of the interval it's dealing with (ie, the month containing
the earlier of the two dates). This is just wrong, because interval
addition adds months first and then days. The appropriate conversion
to use is actually the length of the next-to-last month of the interval.

As an example, 8.2 and CVS HEAD produce

regression=# select age('2007-03-14', '2007-02-15');
age
---------
27 days
(1 row)

which is reasonable, but

regression=# select age('2007-04-14', '2007-02-15');
age
---------------
1 mon 27 days
(1 row)

is not so reasonable, nor is

regression=# select age('2007-03-14', '2007-01-15');
age
---------------
1 mon 30 days
(1 row)

If we change the code to use the next-to-last month of the interval
then these two cases produce '1 mon 30 days' and '1 mon 27 days'
respectively.

Another problem is that the code isn't doing the propagate-to-next-field
bit for negative fractional seconds. Hence it produces

regression=# select age('2007-02-14 01:00:00', '2007-01-15 01:00:00.4');
age
----------------------
30 days -00:00:00.40
(1 row)

which is maybe not incorrect, but certainly fairly inconsistent with

regression=# select age('2007-02-14 01:00:00', '2007-01-15 01:00:01');
age
------------------
29 days 23:59:59
(1 row)

Hence I propose the attached patch. This does not change any existing
regression test outputs, but it does change the example given in the
documentation: age(timestamp '2001-04-10', timestamp '1957-06-13')
will now produce '43 years 9 mons 28 days' not 27 days. Which actually
is correct if you try to add back the result to timestamp '1957-06-13'.
It also appears to fix Palle's example:

regression=# select column1, age(column1, '2006-11-02'), date '2006-11-02' +
age(column1, '2006-11-02') from (values ('2007-01-31'::date),
('2007-02-01')) as alias;
column1 | age | ?column?
------------+----------------+---------------------
2007-01-31 | 2 mons 29 days | 2007-01-31 00:00:00
2007-02-01 | 2 mons 30 days | 2007-02-01 00:00:00
(2 rows)

As I said earlier, I'm worried about changing the behavior of a function
that's been around for so long, so I'm disinclined to back-patch this.
But it seems like a reasonable change to make in 8.3. Comments?

regards, tom lane

Content-Description: age.patch

Index: timestamp.c
===================================================================
RCS file: /cvsroot/pgsql/src/backend/utils/adt/timestamp.c,v
retrieving revision 1.179
diff -c -r1.179 timestamp.c
*** timestamp.c	6 Jul 2007 04:15:59 -0000	1.179
--- timestamp.c	8 Jul 2007 19:45:04 -0000
***************
*** 3044,3050 ****
if (timestamp2tm(dt1, NULL, tm1, &fsec1, NULL, NULL) == 0 &&
timestamp2tm(dt2, NULL, tm2, &fsec2, NULL, NULL) == 0)
{
! 		fsec = (fsec1 - fsec2);
tm->tm_sec = tm1->tm_sec - tm2->tm_sec;
tm->tm_min = tm1->tm_min - tm2->tm_min;
tm->tm_hour = tm1->tm_hour - tm2->tm_hour;
--- 3044,3051 ----
if (timestamp2tm(dt1, NULL, tm1, &fsec1, NULL, NULL) == 0 &&
timestamp2tm(dt2, NULL, tm2, &fsec2, NULL, NULL) == 0)
{
! 		/* form the symbolic difference */
! 		fsec = fsec1 - fsec2;
tm->tm_sec = tm1->tm_sec - tm2->tm_sec;
tm->tm_min = tm1->tm_min - tm2->tm_min;
tm->tm_hour = tm1->tm_hour - tm2->tm_hour;
***************
*** 3064,3069 ****
--- 3065,3081 ----
tm->tm_year = -tm->tm_year;
}

+ 		/* propagate any negative fields into the next higher field */
+ 		while (fsec < 0)
+ 		{
+ #ifdef HAVE_INT64_TIMESTAMP
+ 			fsec += USECS_PER_SEC;
+ #else
+ 			fsec += 1.0;
+ #endif
+ 			tm->tm_sec--;
+ 		}
+ 
while (tm->tm_sec < 0)
{
tm->tm_sec += SECS_PER_MINUTE;
***************
*** 3082,3097 ****
tm->tm_mday--;
}

! 		while (tm->tm_mday < 0)
{
if (dt1 < dt2)
{
! 				tm->tm_mday += day_tab[isleap(tm1->tm_year)][tm1->tm_mon - 1];
! 				tm->tm_mon--;
}
else
{
! 				tm->tm_mday += day_tab[isleap(tm2->tm_year)][tm2->tm_mon - 1];
tm->tm_mon--;
}
}
--- 3094,3130 ----
tm->tm_mday--;
}

! 		/*
! 		 * day-to-month conversion is tricky because variable.  For each
! 		 * decrement in tm_mon, we should adjust tm_mday by the length of
! 		 * the next-to-last month(s) of the original time interval.
! 		 * This corresponds to the notion that interval addition will add
! 		 * months first, then days.
! 		 */
! 		if (tm->tm_mday < 0)
{
+ 			int		end_year;
+ 			int		end_mon;
+ 
if (dt1 < dt2)
{
! 				end_year = tm2->tm_year;
! 				end_mon = tm2->tm_mon;
}
else
{
! 				end_year = tm1->tm_year;
! 				end_mon = tm1->tm_mon;
! 			}
! 
! 			while (tm->tm_mday < 0)
! 			{
! 				if (--end_mon <= 0)
! 				{
! 					end_mon = MONTHS_PER_YEAR;
! 					end_year--;
! 				}
! 				tm->tm_mday += day_tab[isleap(end_year)][end_mon - 1];
tm->tm_mon--;
}
}
***************
*** 3158,3163 ****
--- 3191,3197 ----
if (timestamp2tm(dt1, &tz1, tm1, &fsec1, &tzn, NULL) == 0 &&
timestamp2tm(dt2, &tz2, tm2, &fsec2, &tzn, NULL) == 0)
{
+ 		/* form the symbolic difference */
fsec = fsec1 - fsec2;
tm->tm_sec = tm1->tm_sec - tm2->tm_sec;
tm->tm_min = tm1->tm_min - tm2->tm_min;
***************
*** 3178,3183 ****
--- 3212,3228 ----
tm->tm_year = -tm->tm_year;
}

+ 		/* propagate any negative fields into the next higher field */
+ 		while (fsec < 0)
+ 		{
+ #ifdef HAVE_INT64_TIMESTAMP
+ 			fsec += USECS_PER_SEC;
+ #else
+ 			fsec += 1.0;
+ #endif
+ 			tm->tm_sec--;
+ 		}
+ 
while (tm->tm_sec < 0)
{
tm->tm_sec += SECS_PER_MINUTE;
***************
*** 3196,3211 ****
tm->tm_mday--;
}

! 		while (tm->tm_mday < 0)
{
if (dt1 < dt2)
{
! 				tm->tm_mday += day_tab[isleap(tm1->tm_year)][tm1->tm_mon - 1];
! 				tm->tm_mon--;
}
else
{
! 				tm->tm_mday += day_tab[isleap(tm2->tm_year)][tm2->tm_mon - 1];
tm->tm_mon--;
}
}
--- 3241,3277 ----
tm->tm_mday--;
}

! /*
! * day-to-month conversion is tricky because variable. For each
! * decrement in tm_mon, we should adjust tm_mday by the length of
! * the next-to-last month(s) of the original time interval.
! * This corresponds to the notion that interval addition will add
! * months first, then days.
! */
! if (tm->tm_mday < 0)
{
+ int end_year;
+ int end_mon;
+
if (dt1 < dt2)
{
! end_year = tm2->tm_year;
! end_mon = tm2->tm_mon;
}
else
{
! end_year = tm1->tm_year;
! end_mon = tm1->tm_mon;
! }
!
! while (tm->tm_mday < 0)
! {
! if (--end_mon <= 0)
! {
! end_mon = MONTHS_PER_YEAR;
! end_year--;
! }
! tm->tm_mday += day_tab[isleap(end_year)][end_mon - 1];
tm->tm_mon--;
}
}

---------------------------(end of broadcast)---------------------------
TIP 4: Have you searched our list archives?

http://archives.postgresql.org

--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://www.enterprisedb.com

+ If your life is a hard drive, Christ can be your backup. +

I have applied the attached patch that documents the age() behavior,
plus fixes the mismatch sign for seconds by using part of Tom's earlier
patch.

I agree we want to keep the symmetry we have. We can call this item
closed.

---------------------------------------------------------------------------

Tom Lane wrote:

Pelle Johansson <pelle@morth.org> writes:

If you were to use the result for subtracting from the first value,
instead of adding to the second, the conditions are reversed. It's
not really as obvious as I first thought whether there's 2 months and
29 days or 2 months and 30 days between 2006-11-02 and 2007-02-01...

Hmm, that's a really good point; perhaps the original author was
thinking of it in those terms, in which case using the first month of
the interval is indeed sane. (Almost: I believe that the loop can
iterate more than once, and then you need to look to the second month
etc. The code's not doing that, so there's still a corner-case bug,
plus the fsec issue.)

Other than that corner case, it seems the behavior we currently have is
if x > y, age() produces a positive interval such that
x - age(x, y) = y
if x < y, age() produces a negative interval such that
y + age(x, y) = x

Are we satisfied with just documenting that, or do we want to change it,
and if so to what?

As the code currently stands, we have the symmetry property
age(x,y) = - age(y,x)
for all x,y. I don't think we can preserve that if we try to simplify
the relationship to interval addition/subtraction.

Comments?

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 7: You can help support the PostgreSQL project by donating at

http://www.postgresql.org/about/donate

--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://www.enterprisedb.com

+ If your life is a hard drive, Christ can be your backup. +