Inconsistent `w` symbol formatting from joda time

问题:

My team is looking to switch from Joda time to java.time, but we're seeing different behavior in formatting using the same pattern. The issue arises when we're using the week-of-week-year w symbol:

final String dateString = "2016-01-04 00:00:00";
final String inputPattern = "yyyy-MM-dd HH:mm:ss";

// parse the input string using Joda
final org.joda.time.format.DateTimeFormatter jodaInputFormatter = org.joda.time.format.DateTimeFormat.forPattern(inputPattern);
final org.joda.time.DateTime jodaDateTime = jodaInputFormatter.parseDateTime(dateString);

// parse the input string using two different java.time classes
final java.time.format.DateTimeFormatter javaTimeInputFormatter = java.time.format.DateTimeFormatter.ofPattern(inputPattern).withZone(java.time.ZoneOffset.UTC);
final java.time.LocalDateTime localDateTime = java.time.LocalDateTime.parse(dateString, javaTimeInputFormatter);
final java.time.ZonedDateTime zonedDateTime = java.time.ZonedDateTime.parse(dateString, javaTimeInputFormatter);

final String outputPattern = "'week' w - dd/MM/yyyy HH:mm:ss";
final org.joda.time.format.DateTimeFormatter jodaOutputFormatter = org.joda.time.format.DateTimeFormat.forPattern(outputPattern);
final java.time.format.DateTimeFormatter javaTimeOutputFormatter = java.time.format.DateTimeFormatter.ofPattern(outputPattern);

// output: week 1 - 04/01/2016 00:00:00
System.out.println("With joda: " + jodaOutputFormatter.print(jodaDateTime));
// output: week 2 - 04/01/2016 00:00:00
System.out.println("With LocalDateTime: " + javaTimeOutputFormatter.format(localDateTime));
// output: week 2 - 04/01/2016 00:00:00
System.out.println("With ZonedDateTime: " + javaTimeOutputFormatter.format(zonedDateTime));

For some reason, the output from the w symbol is off-by-one across the two implementations.

What is causing this inconsistency? Is the w symbol inconsistently implemented across Joda time and java.time?

回答1:

Well, it is a little bit speculative, but since you told me that your system timezone is EST (-05:00) I assume that you are sitting in US (New York?). And US does not apply ISO-8601-week rules. Weeks start on Sunday, and the first week of the year does not need to contain at least 4 days (even one day is enough to be counted as first week of year).

So let's look at your example date of 4th of January. It is a Monday. The first US-week is from 2016-01-01 until 2016-01-02 (2 days - enough for US). And the second US-week starts on Sunday the 3rd of January, so the fourth of January is in the second week, too.

And now the critical point: java.time (JSR-310) uses a localized week of week-based-year for the pattern symbol w, see also its backport which should have the same code. Code excerpt:

} else if (cur == 'w') {
    if (count > 2) {
        throw new IllegalArgumentException("Too many pattern letters: " + cur);
    }
    appendInternal(new WeekFieldsPrinterParser('w', count));

...

static final class WeekFieldsPrinterParser implements DateTimePrinterParser {
    private final char letter;
    private final int count;

    public WeekFieldsPrinterParser(char letter, int count) {
        this.letter = letter;
        this.count = count;
    }

    @Override
    public boolean print(DateTimePrintContext context, StringBuilder buf) {
        WeekFields weekFields = WeekFields.of(context.getLocale());
        DateTimePrinterParser pp = evaluate(weekFields);
        return pp.print(context, buf);
    }

The use of WeekFields.of(context.getLocale()) for the pattern symbol "w" is evident.

In contrast, Joda-Time only uses ISO-8601-week-definition which let weeks start on Monday and count that week as first week of year which contains at least four days in current calendar year. So the Monday 4th of January is the start of the first week-of-year because the three days before are not enough for ISO-8601 to be counted as week. Those preceding days are instead considered as last week of previous year.

Consequently, Joda-Time displays week 1 for 4th of January while java.time uses the US-week 2.

Solution of your problem is to specify the locale such that the formatter will use ISO-weeks so you get the same result as in Joda-Time. For example, you could choose Locale.UK which also uses English but other week rules. Don't rely on your default locale. This can fool you.

回答2:

Edit: As Richard points out, I'm wrong—Java SE actually does say that the first week of a week-based year is the first Monday-based week containing at least four days, just like Joda-Time.

From the documentation of Java SE's IsoFields.WEEK_OF_WEEK_BASED_YEAR:

The week-of-week-based-year has values from 1 to 52, or 53 if the week-based-year has 53 weeks.

No mention is made of excluding any weeks, so it makes sense to assume that all weeks are counted.

From the Joda-Time Fields overview:

Weeks run from 1 to 52-53 in a week based year. The first day of the week is defined as Monday and given the value 1. The first week of a year is defined as the first week that has at least four days in the year.

January 1 and 2 comprise the first partial week of 2016, and since that's fewer than four days, Joda-Time does not count it as a week at all. January 4 is in the first week which contains four or more days.