[LEAPSECS] Java: ThreeTen/JSR-310

[Gerard Ashton]

When it comes to finding a standard to smooth UTC in order to hide leap 
seconds for purposes of the Java Instant class, I would be tempted to 
find a computer architecture that handles leap seconds properly, and 
propose a standard that can be implemented as easily as possible on that 
architecture. Reward the architecture that does it right and punish the 
architecture that does it wrong.

One architecture that appears to do it right, if you want it to, is the 
IBM mainframe z-series. There is a Redbook that describes how it works 
in great detail:

http://publib-b.boulder.ibm.com/abstracts/sg242070.html?Open

However, the Redbook concentrates on how the hardware clocks are 
affected, and does not explain fully how this is carried through to 
application programs. A further complication appears to be that various 
installations seem to be able to set their computers to keep track of 
leap seconds or not. Any timing error, whether a leap second when leap 
second handling is turned off, or from a different source, can be 
resolved by steering the system clock. Steering will fix a 1 second 
error in about 12 hours. Steering is invisible to the operating system 
and application programs.

So the IBM approach might be too flexible for Java; it would probably be 
impractical to allow Java to work differently at different installations.

Windows and Unix have general reputations of not doing it right; does 
anyone know of a hardware/operating system combination that handles leap 
seconds correctly? If so, does it have a defined approach to providing a 
quasi-UTC that hides leap seconds to applications that wish to live in 
blissful ignorance?

Gerry Ashton