News for hackers. Stuff that matters.
Re:Text Editing Hates You Too by demurgos on Tuesday October 29, 2019 @10:42AM (#21387678)
> I always convert time into time in seconds since the epoch

You should still be careful how you define a _second_. A POSIX second is not the same as an SI second. A POSIX second corresponds to 1/(24 * 60 * 60) of a day. An SI second is defined using atomic time measurement. It means that the duration of a POSIX second depends on the length of the day, and some days are longer than others due to leap seconds.

The worst part is that most tools rely on POSIX timestamps instead of TAI timestamps (based on the SI definition). One of the consequences of using POSIX timestamps is that taking the difference between two timestamp does not safely return the number of elapsed seconds.

Re:Text Editing Hates You Too by haberman on Tuesday October 29, 2019 @12:15PM (#21388827)
> A POSIX second corresponds to 1/(24 * 60 * 60) of a day. An SI second is defined using atomic time measurement.

I don't think that is quite accurate. Traditionally Unix Time seconds are SI seconds, that tick in sync with UTC and atomic clocks (modulo clock error). To account for leap seconds, Unix Time has traditionally been discontinuous when a leap second occurs. Each second that ticks is still an SI second, but the sequence of Unix Time around a leap second will be X-2, X-1, X, X, X+1, X+2, etc.

This discontinuity means that every Unix Time day has exactly 86,400 seconds cumulatively, even though some days tick 86,401 elapsed SI seconds. So it is true that computing a difference between two Unix Time values does not give you a reliable measure of how many SI seconds have elapsed, as you say. But the individual seconds are still all SI seconds.

More recently it has become popular to smear the leap second instead. Google pioneered this approach in 2008 (https://googleblog.blogspot.com/2011/09/time-technology-and-...) and now advocates making it standard (https://developers.google.com/time/smear). This approach does vary the actual length of a second during the 24 hours surrounding a leap second. But this doesn't come from POSIX, it's a departure from previous practice that is designed to avoid a discontinuity.

Re:Text Editing Hates You Too by demurgos on Tuesday October 29, 2019 @07:51PM (#21393335)
Yes, I definitely agree that the reality is even more complicated and there are different implementations. The recent time-smearing approach is closer to the POSIX standard [0]:

> How any changes to the value of seconds since the Epoch are made to align to a desired relationship with the current actual time is implementation-defined. As represented in seconds since the Epoch, each and every day shall be accounted for by exactly 86400 seconds.

Another fact that I find interesting is that the use of discontinuities or time-smearing of the leap days are actually modern approaches. Nowadays, POSIX and TAI differ by an integer amount of seconds on regular days, and a fractional amount when smearing the leap-second day. At the beginning, the leap second duration between POSIX and TAI was smeared over years from one leap second to the other. So the difference between both timestamps would change continually other these periods!

My main takeaway is that POSIX timestamps are based on calendar days while TAI timestamps are based on physics constants. POSIX timestamps are a tradeoff: they're not suitable for time computations (durations, or even uniquely identifying an instant when discontinuities are involved) but they don't require a leap seconds lookup-table to display time for humans... in UTC... Which means you still need lookup tables for timezones, countries, DST, calendar changes, etc. Basically my opinion is that POSIX time is neither the best to precisely keep track of time nor to format it for humans. It just sits in between these two use cases because it was good enough at the time and then too hard to change.

[0]: https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1...

Re:Text Editing Hates You Too by haberman on Tuesday October 29, 2019 @08:40PM (#21393581)
> At the beginning, the leap second duration between POSIX and TAI was smeared over years from one leap second to the other. So the difference between both timestamps would change continually other these periods!

Do you have any reference to a system that was implemented this way? I have never heard of such a thing and I don't see how it could be implemented.

The insertion of leap seconds is unpredictable and occurs when the International Earth Rotation and Reference Systems Service determines that the difference between UTC and UT1 are approaching 0.9 seconds. Leap seconds are generally only decided six months or so before they occur. So you don't know at the beginning of the year how many leap seconds the year will have. So it doesn't seem possible to smear over a year.

Re:Text Editing Hates You Too by demurgos on Wednesday October 30, 2019 @05:20AM (#21395827)
I remember reading it from Wikipedia some times ago. I looked it up again and I was wrong: this was not applied for POSIX but UTC itself. The arrival of international networks seems to actually be the reason why they stopped smearing it continually and switched to a linear version with discontinuities. [0]

> UTC is a discontinuous time scale. From its beginning in 1961 through December 1971 the adjustments were made regularly in fractional leap seconds so that UTC approximated UT2. Afterwards these adjustments were made only in whole seconds to approximate UT1. This was a compromise arrangement in order to enable a publicly broadcast time scale; the post-1971 more linear transformation of the BIH's atomic time meant that the time scale would be more stable and easier to synchronize internationally.

Wikipedia sources it to a book [1].

[0]: https://en.wikipedia.org/wiki/International_Atomic_Time#Rela...

[1]: McCarthy, Dennis D.; Seidelmann, Kenneth P. (2009). Time: From Earth Rotation to Atomic Physics. Weinheim: Wiley-VCH Verlag GmbH & Co. KGaA. ISBN 978-3-527-40780-4.