FWD » Core Development » Base Language

The issue comes from inconsistent calendar handling during date conversions, not from any specific date. Different parts of the code rely on Java’s GregorianCalendar, which implicitly mixes Julian and Gregorian rules depending on how it’s initialized, leading to different millisecond values for the same logical date and, after persistence, a one-day shift.

A key point here is that Java’s modern date/time model follows the Proleptic Gregorian calendar, which applies Gregorian rules backward in time to keep calculations consistent. This differs from historical reality, where the Julian calendar was used before the Gregorian calendar reform, and the gap between the two systems varies over time. Because our code paths are not consistently aligned with a single calendar system, conversions can yield slightly different results, which then surface as off by a number of days errors.

I suggest we revisit how we use the Calendar, Date, and OffsetDateTime APIs and consider moving to a more consistent approach, or at least ensuring they are used in a uniform way across the codebase.

• the 1-day difference for 'more recent dates' is, in my current opinion, caused by the date arithmetic involving timezone component while intermediary conversion to datetime(-tz) values. I think this complicates the computations, also affecting performance;
• r16540 (11370b) addressed only the 0-year component of proleptic Gregorian calendar. The date skew need also be fixed. Ancient (pre oct' 1582) dates are incorrectly converted to OffsetDateTime (the intermediary native date type used low-level in BaseRecord.data[]) and eventually persisted to databases;
• date format issue only for negative years (in date.getIsoDate()) causing to display them as 000-100 instead of -100.

Ovidiu, I committed the first round of fixes to 11421a. I still have to address the getIsoDate issue. If you have a few minutes, please look in the branch and confirm that the fixes are right, thanks.

In date.java:4579: the line seems useless. The assignment of time field (setTimeInMillis()) is overwritten a few lines below (at 4588), after millis is slidden to UTC.

• apply correction only if the offsetDays is 0 and possibly stick to original code if the date is recent (year > 1582). I am not aware how costly are the new time-related operations (I guess they require validations and other non-trivial arithmetic), but usually we will be processing values from latest centuries, and any computation shortcut is useful. In other words, let's spend extra CPU cycles strictly on exceptions, not on normal dates.
• the thresholds checks from julianToGregorianOffsetDays() can be simplified. For example: && !d.isAfter(LocalDate.of(1582, 10, 14) from the second condition can be dropped since that was the exit condition from previous if.

Ovidiu Maxiniuc wrote:

In date.java:4579: the line seems useless. The assignment of time field (setTimeInMillis()) is overwritten a few lines below (at 4588), after millis is slidden to UTC.

I know it seems like it, but this solved a corner case, the calendar is a mutable, reused variable, something about setting the time prior to the other operations made it behave correctly.
This is the test case it solved

define temp-table tt1 fields dt as date.
DEFINE VARIABLE my-date AS DATE NO-UNDO.

my-date = DATE("02/28/1900").

create tt1.
tt1.dt = my-date.
release tt1.
find first tt1.
message tt1.dt = my-date. // FWD -> false

It prints yes for me in both cases (with and without the extra setTimeInMillis()).

But there is something that does not look fine. Inspecting the tt1 buffer with my IDE the date is, indeed, {java.sql.Date@13275} "1900-02-28". But when expanded, it is odd: it has the time component set to 1:44:24 only to balance it out with +144 timezone:

result = {com.goldencode.p2j.persist.TemporaryBuffer@13189} tt1: 2304 [null, null, null, null, null, null, 1900-02-28]RecordState{CACHED }/1
└ currentRecord = {com.goldencode.work.dmo._temp.Tt1_1_1__Impl__@13190} tt1: 2304 [null, null, null, null, null, null, 1900-02-28]RecordState{CACHED }
  └ data = {java.lang.Object[7]@13194} 
    └ 6 = {java.sql.Date@13275} "1900-02-28" 
      ├ cdate = {sun.util.calendar.Gregorian$Date@14479} "1900-02-28T01:44:24.000+0144" 
      └ fastTime = -2203977600000

OTOH, evaluating new java.sql.Date(0, 1, 28) results in:

result = {java.sql.Date@15334} "1900-02-28" 
├ cdate = {sun.util.calendar.Gregorian$Date@15336} "1900-02-28T00:00:00.000+0144" 
└ fastTime = -2203983864000