[Java][FlightSQL] JDBC driver returns pre-1970 dates one day late when a Calendar is supplied

Commencez par DateTimeUtils.getTimestampValue(long), puis examinez ArrowFlightJdbcDateVectorAccessor.getDate(Calendar) et le test existant pour les dates négatives. Reproduisez le cas 1950-06-01 01:00:00 UTC avec un décalage de calendrier et vérifiez que les calculs du jour de l’époque et de l’heure utilisent des limites de jour correspondantes. C’est terminé lorsque les dates antérieures à 1970 avec des calendriers fournis renvoient la date correcte et que le test de régression réussit.

Describe the bug, including details regarding any error messages, version, and platform.

DateTimeUtils.getTimestampValue(long) splits epoch milliseconds into an epoch day and a time within that day. The old code used / and %:

public static Timestamp getTimestampValue(long millisWithCalendar) {
  long milliseconds = millisWithCalendar;
  if (milliseconds < 0) {
    // LocalTime#ofNanoDay only accepts positive values
    milliseconds -= ((milliseconds / MILLIS_PER_DAY) - 1) * MILLIS_PER_DAY;
  }

  return Timestamp.valueOf(
      LocalDateTime.of(
          LocalDate.ofEpochDay(millisWithCalendar / MILLIS_PER_DAY),
          LocalTime.ofNanoOfDay(TimeUnit.MILLISECONDS.toNanos(milliseconds % MILLIS_PER_DAY))));
}

Both operators round toward zero. For dates before 1970, the code adjusted the negative remainder because LocalTime.ofNanoOfDay rejects it. It did not adjust the epoch day. When the input was negative and not exactly midnight, the day and remainder then referred to different days. The date came back one day late.

For example, -618102000000 ms is 1950-06-01 01:00:00 UTC. Division by 86400000 truncates to -7153, which is 1950-06-02. The remainder is 3600000 ms, or 01:00. The old result was 1950-06-02 01:00:00. Math.floorDiv returns -7154, the correct epoch day for 1950-06-01.

How this happens in JDBC

ArrowFlightJdbcDateVectorAccessor.getDate(Calendar) applies the calendar offset before calling this method. A DATE starts at midnight, but any difference between the supplied calendar zone and the JVM default moves it away from midnight. Passing a Calendar is the documented way to read a date in a specific zone. As a result, every pre-1970 date read this way came back one day late. Any non-zero offset can expose the bug. Historical offsets can include seconds, such as +05:53:28.

The existing negative test used -618105600000, exactly 1950-06-01 00:00:00 UTC. At midnight, truncating division and floor division agree, so that test could not expose the bug.

The fix uses Math.floorDiv and Math.floorMod for both parts and removes the manual adjustment. Positive values behave as before. A test now covers 1950-06-01 01:00:00 UTC.

This affects flight-sql-jdbc-core on main and in 19.0.0. It is platform independent. Issues #732 and #324 cover broader timestamp and time-zone behavior. This report is only about the integer-division bug above.