A general thanks to aphyr for exploring all of these kinds of issues in distributed datastores at a level where someone without a lot of database knowledge can understand and reason through. If anyone hasn't read his Jepsen series and is interested in these kinds of things, it is well worth a read.
Of course, I wouldn't read it if you store very important data in any of the datastores he talks about, you might be scared to learn how your system actually operates =).
P.S. - The Spanner link is broken, you have some invisible character being added to the end, gets encoded to %E2%80%8E
A little surprising that this article does not mention HBase which uses timestamps as part of the persisted key, while using an entirely different concept of a "WriteNumber" as an internal ID to resolve conflicting updates (MVCC).
WriteNumbers are a logical clock provided by (last I checked) a single coordinating HBase node, which is a totally valid and commonly used way to order operations. For the purposes of this article, though, I'm trying to focus on LWW data models and wall clocks. A discussion of every database's approach to ordering would take a little longer. ;-)
So last time the leap second issue came up, I learned that it's an "issue" because (as this article mentions) POSIX defines one day as exactly 86400 seconds. That's clearly incorrect; why do we still want to keep that definition around?
Well... that page refers several times to an apparently heated dispute over whether "day" shall be defined solely with reference to cesium atoms, or whether "day" shall be defined with reference to the rotation of the earth. It notes the clash noted in the article between wanting to know time elapsed vs. wanting to know current absolute time.
I don't see that that speaks particularly to my question. That clash only exists because of a background assumption that a day must necessarily consist of 86400 seconds with unique names. Then, when an 86401-second day comes along, the elapsed-time people say we should give them all unique names, and the absolute-time people appear to believe that that would be a catastrophic failure of the principle that there are 86400 of them. But that's not actually a catastrophe. It's easy to determine absolute time given time elapsed from another absolute time.
We could analogously ask whether "year" should be defined with reference to the orbit of the earth around the sun, or solely by reference to the rotation of the earth, and in fact this was a historical dispute the wrong end of which persists today in the russian orthodox calendar. But we all know the russian orthodox calendar is wrong. It wasn't difficult to say that a year is usually 365 days but sometimes 366. It's also not difficult to say that a day is usually 86400 seconds but sometimes 86401... and all the technical arguments for why that shouldn't be the case seem (to me) to be equally applicable to the variable year "problem". What's the value of keeping the POSIX standard out of touch with reality, just because it was originally written to be out of touch with reality?
The state of the art now seems to be Google's system for complying with POSIX by lengthening the duration of a second on days which contain leap seconds. That also used to be standard practice; the day was twelve hours long by definition, and the night was twelve (different) hours long by definition, and day-hours and night-hours varied in length with the seasons. Should we go back to that system? Why is it a good idea for seconds? The entire point of defining seconds in terms of cesium was to stop defining them as 1/86400 of a day. But if you stop defining them as 1/86400 of a day, how does it come as a surprise that a day might contain other than 86400 of them?
Why does everyone refer to this as "Google's system"? The idea of slowing down clocks instead of turning them back is very-very old. In fact 10 years ago one of the first thing we were thought at university in the Embedded Systems course is that you never turn clocks back. Am I missing something here? (honest question, no irony or sarcasm intended)
This is not slowing down a clock for a while because it runs fast, but slowing it down to make it run too slow, and then skipping a leap second.
The first is good; it corrects an error without introducing the catastrophe of a clock running backwards. The second, according to some is bad because it introduces a temporary error without any benefit. According to others, it does introduce a benefit: you won't have to deal with leap seconds.
> So last time the leap second issue came up, I learned that it's an "issue" because (as this article mentions) POSIX defines one day as exactly 86400 seconds. That's clearly incorrect; why do we still want to keep that definition around?
If we didn't do that it would be impossible to express times more than about 6 months into the future as POSIX timestamps, since we wouldn't know how many leap seconds there would be between now and then.
The point of 86400 is to give a fixed way to calculate the date from gettimeofday(2), allowing people to write code that refers to future dates in the same way as past dates. Using variable-length days makes future dates indeterminate, which would cause no end of mischief for interest rate and billing rate calculations. (Think fraud—allocate the budget for one second before a potential leap second, pay the expenses after the leap second has been added to the table.)
Of course, I wouldn't read it if you store very important data in any of the datastores he talks about, you might be scared to learn how your system actually operates =).
P.S. - The Spanner link is broken, you have some invisible character being added to the end, gets encoded to %E2%80%8E