time: unable to parse UTC leap seconds
aecolley opened this issue · comments
What does 'go version' print? go version go1.3.1 linux/amd64 What steps reproduce the problem? Attempting to parse any valid leap second, such as "2005-12-31T23:59:60Z". http://play.golang.org/p/lBYHT8mwbb What happened? parse error: parsing time "2005-12-31T23:59:60Z": second out of range What should have happened instead? parsed as: 2005-12-31T23:59:60Z Please provide any additional information below. http://golang.org/src/pkg/time/format.go has the immediate bug at lines 818-822: case stdSecond, stdZeroSecond: sec, value, err = getnum(value, std == stdZeroSecond) if sec < 0 || 60 <= sec { rangeErrString = "second" } It seems that there's a fundamental restriction in the time package, where it repeats the expensive POSIX mistake of assuming that you can ignore leap seconds if you're not doing high-precision radio astronomy. Since leap seconds don't go away just because you can't represent them in the most convenient time datatype, this makes the time package useless for programs that are supposed to continue to run without crashing during a leap second. Suggested fixes: 1. Full support for conversion between monotonic counter (TAI) <=> wall time (UTC) <=> POSIX time_t. The API is already flexible enough for this, but the implementation would need to become aware of leap seconds. 2. Change the package documentation so it gives fair warning that "UTC" doesn't mean real UTC, but has undefined behavior for leap seconds. 3. Abolish leap seconds. There may be politics involved. 4. Declare this behavior "working as intended". Make it someone else's problem to cope with the next leap-second production crisis (scheduled for 2015). Deplore the horrible hacks they perpetrate to keep the site up, eg: http://googleblog.blogspot.com/2011/09/time-technology-and-leaping-seconds.html
Has this been kicked off into the long grass?
It's a shame it isn't getting some attention (especially in view of Java8 having finally sorted out the Java implementation https://docs.oracle.com/javase/8/docs/api/java/time/package-summary.html).
This is somewhat related to Cloudflare's outtage this year: How and why the leap second affected Cloudflare DNS
It would not be difficult to support leap seconds. See https://commons.wikimedia.org/wiki/File:Avoid_Using_POSIX_time_t_for_Telling_Time.pdf for some suggestions on how.
@PaluMacil - Not really. There was no parsing of ":60" involved in that particular problem.
Note that not very many other languages support parsing :60
in the seconds field. For example, JavaScript returns an Invalid Date
, and .NET throws a FormatException
. I'm sure others give similar errors.
Those that do stub it for :59
, such as Java 8's new time API, which describes it's handling of leap seconds here: https://docs.oracle.com/javase/8/docs/api/java/time/format/DateTimeFormatter.html#parsedLeapSecond--
Ultimately, the problem is that most internal data structures store a number of units since some epoch, and don't track a table of leap seconds. Doing so would come at a performance and maintenance cost that's usually not worth it.
Unix Time, for example, is explicitly the number of non-leap seconds since 1970-01-01T00:00:00Z
.
Also note that #15247 already documented that leap seconds are not accounted for in calculations. I see no reason that parsing should be any different.
The options here are either (1) parse seconds=60 as seconds=59, or (2) continue to return an error. Given that we've returned an error this far with minimal fallout and also that .NET and even JavaScript (not exactly known as a picky language!) return errors, I am comfortable continuing to return an error.