[LEAPSECS] leap smear

[Michael Deckers]

    On 2011-09-16 22:46, Daniel R. Tobias observed
    Google's lack of compliance with international standards:


>  On 16 Sep 2011 at 10:54, Poul-Henning Kamp wrote:



>  >  Well, Googles hack is a workaround for internal use with no regard

>  >  to subsecond interoperability with anybody else.



>  That's typical of Google's general philosophy, where they value their

>  services functioning smoothly over strict compliance with external

>  standards; for instance, "web purist geeks' find it infuriating that

>  Google's HTML fails to validate, but they're more interested in

>  shaving off a few bytes by doing things like omitting quotes around

>  attributes even where the standards require them, so long as all

>  known browsers render the page correctly.


    There simply is no standard modification of UTC that
    is a monotone function of TAI, and which is N times
    continuously differentiable for some N >= 0. Instead, there
    are several such modifications, used or propsed, such as
    Google's, the SLS time scale of Markus Kuhn, the
    modification in SOFA, and the time_t values in UNIX-like
    operating systems, each especially designed for their
    usage.

    Google needed a modification of UTC that could be used as
    the steering input for an ensemble of clocks. This requires
    a signal that is at least two times differentiable, like
    the signal of a clock that is not subject to sudden
    changes in rate. I find it clever from the Google
    engineers that they realized that the smoothness of
    the steering clock must be controlled, and that the
    the maximal deviation in rate is a secondary concern
    that is best left to a parameter of the solution.

    As long as that window size is less than the distance
    between successive leap seconds, UTC[Google] is well-defined
    -- eg, when UTC was 2008-12-31T23:59:60 (beginning of a
    "positive leap second"), UTC[Google] was 2008-12-31T23:59:59.
    I have never seen an equally clear statement for
    the value of type time_t to be returned by POSIX calls
    of time(), even with parameters left unspecified.

    Michael Deckers.