Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

UTC is good but not enough; it may still bite you in the behind.

What's the time difference between these 2 timestamps:

    2016-12-31 23:59:50 UTC    (unix time 1483232390)
    2017-01-01 00:00:10 UTC    (unix time 1483232410)
That's right: 21 seconds, due to the added leap second 23:59:60.

So then your next option is to use TAI (international atomic time). Of course for most applications, this extra/missing leap second doesn't matter.

And that's the crux of the issue. Given a level of abstraction, we're comfortable with a sufficient level of accuracy because it's the most convenient for 99% of the use cases. It takes a widely adopted tooling to bump the abstraction to a more accurate representation. In UNIX land, going from YYYYMMDDHHMMSS to a number of seconds has been a huge improvement. It's comfortable to use because all environments have the tooling to go back and forth from seconds to human-readable dates.

Try to use TAI and you start feeling a lot more lonely (ever tried to read logs timestamped in TAI? fun!) Most OS and server processes are used with the tradeoff that it's not a big deal if we have 2 events spaced a second apart and both timestamped the same (1483232399), since it's not that frequent.



I wrote a TAI parser library for Go (tai64/tai64n), and I agree that TAI timestamps are... terse..

Nice format though, other than readability. ;)


Wait: are you saying that Unix time does not take into account the leap second?!


Correct. POSIX time is defined as number of seconds per (non-leapsecond) day times number of complete day since the epoch, plus the number of seconds in the current day.

POSIX timestamps get weird around leap seconds. Time stamps get repeated, subtraction returns incorrect results, and so on.


If you use Java and the java.time API they added in the last release, you're covered for all of this.

https://docs.oracle.com/javase/8/docs/api/java/time/Instant....

In this segment, the Java Time-Scale is identical to UTC-SLS. This is identical to UTC on days that do not have a leap second. On days that do have a leap second, the leap second is spread equally over the last 1000 seconds of the day, maintaining the appearance of exactly 86400 seconds per day.

Java apps don't use UNIX time, they use something very similar called the Java timeline which is basically the same as UTC but leap seconds are smeared.

The java.time API is really incredibly thorough. It does of course have support for Japanese dates:

https://docs.oracle.com/javase/8/docs/api/java/time/chrono/J...

The type system and runtime checks also catch subtle errors like assuming "one day from now" and "in 24 hours from now" are the same thing.


IIRC google and the like solve this using a special time routine in their kernels that literally just spread the second out across the day. It's close enough.. and legal will still accept it for auditing requirements.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: