["Matthias Urlichs" <smurf at work.smurf.noris.de>] POSIX vs "right" time difference

Paul Eggert eggert at twinsun.com
Fri May 23 18:53:38 UTC 1997

   Date: Thu, 22 May 1997 15:06:40 +0200 (Funky)
   From: "Matthias Urlichs" <smurf at work.smurf.noris.de>

   the POSIX time zone files ignore leap seconds, i.e. are based
   on GMT, while the 'right' timezone take leap seconds into account.

   The base difference between those, according to "date", is 20 seconds (soon
   to be 21). That's wrong. The difference between UTC and TAI ("atomic time")
   is actually 30 (soon to be 31) seconds; the 10 seconds come from the fact
   that UTC and TAI were initially synchronized in 1958, and in 1972 it was
   realized that they've drifted apart... instant 10-second insertion.

That's not what happened.  There was no 10-second insertion in 1972.
The atomic timekeepers didn't wake up from a 14-year sleep and notice
that the world's clocks had drifted!

Here's a more accurate description of what happened.  Before 1972, UTC
was adjusted continuously, without using leap seconds.  At the moment
of the POSIX Epoch (1970-01-01 00:00:00 UTC), UTC and TAI differed,
not by an integral number of seconds, but by a fractional number.  The
difference between UTC and TAI did not become an integer until 1972,
when the leap second regime was introduced.

Therefore, it is impossible, on a system with integral time_t, to
simultaneously support both "right" and POSIX times accurately before 1972.

In some sense, the "right" approach yields undefined times before
1972, since the "right" approach assumes leap seconds.  But we could
attempt to extend the "right" approach by guessing what it would have
been like had it existed.  That is, with an integral time_t, we could
_approximate_ "right" times before 1972 by consulting historical
tables of the difference between UTC and TAI, and rounding to the
nearest integer.  This is unsatisfactory (not least because such
tables understandably do not go back to times before atomic clocks
were introduced!), but it is probably the best that we can do with the
"right" approach assuming integral time_t.

I view this improvement as a low-priority task since very few people
use the "right" approach these days; the people who care about leap
seconds are using other interfaces.  Furthermore, the philosophy of
the tz package is to describe times as they actually appeared at the
time; it's hard to combine this with the notion of gedanken atomic
clocks ticking away in, say, 1901.

More information about the tz mailing list