[LEAPSECS] Java JSR-310 TAIInstant class

[Michael Deckers]

    On 2011-01-29 22:21, Stephen Colebourne wrote:


>  What JSR-310 needs is a single time-scale that has a single

>  incrementing number of fixed size units (SI seconds) with no other

>  interpretations (such as what a day is). I believe that TAI best meets

>  that definition today, and thus is the choice made.


    IMHO, a programming language should
    *  recognize the fact that there are many time scales, none
       of which is better or worse than the others, and for which
       any given implementation may or may not provide real time
       clocks;
    *  support the relationships between time scales as they occur
       in practice (eg, UTC -- TAI with lepa seconds (and milliseconds),
       UTC -- time zone times, order relation of UTC timestamps,
       etc).

    But it should not:
    *  make unfounded assumptions on time scales (like any time
       scale can be reliably converted to TAI and back -- TAI exists
       only since 1958 but datetimes cover a much broader range);
    *  expose any internal constructs (like auxiliary time scales).

    The goal should be to model reality as far as possible and
    desirable (eg, one might decide to ignore archeological time
    scales at first). Squeezing reality to a model that does not
    quite fit just to make things simpler is, in the long run, bad
    for users and designers. (This is the essential theme of
    this list, by the way.)

    I wouldn't comment on this if JAVA were not such an important
    language. With your UTC-SLS time scale, for example, you prohibit
    correct comparison of timestamps with nanosecond precision during
    positive leap seconds. This is the typical design inconsistency
    that can cause implementation errors and, ulitmately, viruses.

    Michael Deckers.