Thank you for your feedback Weston and Antonie. I agree that ordering discussion should be out of scope for the Arrow format spec. I have removed reference of ordering in the PR so now the only change is mentioning leap seconds to keep it consistent with other temporal types.
I would like to add that even though we are not explicitly discussing ordering in the spec, any kind of restriction we assign to a type would still implicitly impact ordering in downstream compute kernels. This is why I also took out the discussion of leap days in my PR as well. Thanks, QP On Tue, Sep 14, 2021 at 12:46 AM Antoine Pitrou <anto...@python.org> wrote: > > > I agree with Weston that ordering isn't in the scope for the Arrow > format spec (*). For example, implementations are free to define UTF8 > comparisons and ordering as they wish (some may want to invest in the > complexity of the official Unicode collation algorithm, others may be > content with a simple codepoint-wise lexicographic comparison). It > doesn't prevent them from exchanging UTF8 data unambiguously using Arrow. > > (*) It may be in the scope for a hypothetical Compute IR spec, however. > > Regards > > Antoine. > > > Le 14/09/2021 à 07:16, QP Hou a écrit : > > Good point Weston. My proposal was written with the impression that > > Arrow does want to define semantic for some of these temporal types > > based on the existing comments in the Schema.fbs file. > > > > For example, here is a quote taken from the comments for the Time time: > > > > /// This definition doesn't allow for leap seconds. Time values from > > /// measurements with leap seconds will need to be corrected when ingesting > > /// into Arrow (for example by replacing the value 86400 with 86399). > > > > Here is another quote for the Date type: > > > > /// * Milliseconds (64 bits) indicating UNIX time elapsed since the epoch > > (no > > /// leap seconds), where the values are evenly divisible by 86400000 > > > > For the interval type, we have: > > > > // A "calendar" interval which models types that don't necessarily > > // have a precise duration without the context of a base timestamp (e.g. > > // days can differ in length during day light savings time transitions). > > > > I think pushing the responsibility to define these semantics to the > > data producer side is also a perfectly fine design with its own > > trade-offs. It would make data exchange between two different systems > > a little bit harder because consumers need to be aware of the > > semantics defined by the producer. On the other hand, it does make the > > producer implementation easier. It also makes data exchange within the > > same system more efficient if that system's temporal type semantic is > > different from what's defined in Arrow's spec. > > > > Either way, I think it would be good if we can be consistent on our > > temporal type semantics in the spec. If we are making the claim that > > leap seconds should not be taken into account for Time, Timestamp and > > Date types, then it seems natural to make this claim for Interval type > > as well. Alternatively, we could update the spec to make all temporal > > types leap seconds agnostics. > > > > On Mon, Sep 13, 2021 at 12:03 PM Weston Pace <weston.p...@gmail.com> wrote: > >> > >> One could define a sorting based on 30 days months, 365 day years, and > >> 24 hour days. It would be consistent but can lead to some surprising > >> results. It appears that this is what postgres does as I got the > >> following ordering for an interval: > >> > >> 359 days, 12 months, 360 days, 1 year, 365 days, 366 days > >> > >> On the other hand, Joda time forbids comparison of periods (their > >> version of what we call an interval) and offers three ways to convert > >> to a duration. There is toDurationFrom(instant), > >> toDurationTo(instant) which give durations from specific calendar > >> ranges and then there is toStandardDuration() which converts to a > >> duration based on 24 hour days. However, toStandardDuration will > >> still fail if the period has >0 months or years (presumably because > >> months and years are too inconsistent). > >> > >> I'm not sure though that this is something that Arrow needs to define. > >> We aren't specifying any invalid ranges of values. I don't foresee > >> any interoperability concerns. A system that treated intervals as > >> comparable (and didn't factor in DST, leap years, etc.) will read and > >> write intervals the same way as a system that considers intervals > >> incomparable. > >> > >> This question seems to fall into the "compute" space inhabited by > >> topics like "is 'false && null' a false value or a null value" and > >> "should addition overflow or throw an exception". > >> > >> On Mon, Sep 13, 2021 at 6:23 AM QP Hou <houqp....@gmail.com> wrote: > >>> > >>> On Mon, Sep 13, 2021 at 6:18 AM Antoine Pitrou <anto...@python.org> wrote: > >>>> The Duration type is defined with a TimeUnit. You are probably thinking > >>>> about the Interval type. > >>>> > >>> > >>> Oops, my bad, yes, it should be Interval type not Duration. > >>> > >>>> Ok. How about daylight savings? I suppose they are taken into account > >>>> as well. > >>>> > >>> > >>> Yes, the day component in both DAY_TIME and MONTH_DAY_NANO all take > >>> into account of daylight savings.