Re: Incorrect leap seconds in runit's log timestamps

["Joël Riou" <joel.riou@normalesup.org>]

Le mercredi 19 janvier 2005, Lloyd Zusman a écrit :
> I believe that TAI and UTC are 21 seconds apart these days, although a
> while ago, the discrepancy was only 11 seconds.

rant {

No, that is not true. First, a very good link on TAI/UTC questions is 
the following page by David A. Madore :
<URL: http://www.eleves.ens.fr:8080/home/madore/misc/time.html >.

Nowadays, the "difference" between TAI and UTC is *precisely* 32 seconds
(see <URL: http://hpiers.obspm.fr/iers/bul/bulc/bulletinc.dat >).

But, for some obscure reason, the epoch is considered to be 
1970-01-01 00:00:10 TAI on Unix systems whose clock is supposed to
correspond to TAI timestamps (i.e. the time() function will return the
number of seconds since that epoch (*), whereas with a GMT-time, this
function is supposed to return the number of non-leap seconds since the
epoch (**)). 32-10=22, so the difference between the result of time() in
system with TAI/UTC-times seems to be *22*.

(*) It is the case with right/foo (e.g right/Europe/Paris) glibc timezones.

(**) Due to the fact that the difference between TAI & UTC was not an
integer number of seconds in the beginning of the '70, the epoch
considered by TAI and UTC systems do not match exactly [so why the hell
1970-01-01 00:00:10 TAI was chosen as an origin instead of
1970-01-01 00:00:00 TAI ?!]

To be more specific, provided I understand well both runit and skalibs
tai code, the tai functions assume that time()/gettimeofday() return the
number of seconds since 1970-01-01 00:00:10 TAI, which I think is the
Right Thing ; so if your computer uses another convention (UTC ?), you
can get confusing results. [at this point, my brain is too damaged to
check that this precisely explains your particular trouble...]

}

More seriously, the main trouble in this domain is that, as far as I
know, ntpd assumes that the Unix time is ruled by UTC [which is quite
confusing because the NTP protocol has allegedly some stuff to propagate
the table of leap seconds from servers to clients]. I used to think about
some dynamic library to LD_PRELOAD to ntpd to emulate UTC time()
functions on a system with TAI time, but that would be very ugly.

-- 
Joël Riou