Re: Infinite Interval

On Sat, Dec 17, 2022 at 2:34 PM Joseph Koshakow <koshy44@gmail.com> wrote:
>
> Hi Ashutosh,
>
> I've added tests for all the operators and functions involving
> intervals and what I think the expected behaviors to be. The
> formatting might be slightly off and I've left the contents of the
> error messages as TODOs. Hopefully it's a good reference for the
> implementation.
>
> > Adding infinite interval to an infinite timestamp with opposite
> > direction is not going to yield 0 but some infinity. Since we are adding
> > interval to the timestamp the resultant timestamp is an infinity
> > preserving the direction.
>
> I think I disagree with this. Tom Lane in one of the previous threads
> said:
> > tl;dr: we should model it after the behavior of IEEE float infinities,
> > except we'll want to throw errors where those produce NaNs.
> and I agree with this opinion. I believe that means that adding an
> infinite interval to an infinite timestamp with opposite directions
> should yield an error instead of some infinity. Since with floats this
> would yield a NaN.
>
> > Dividing infinite interval by finite number keeps it infinite.
> > TODO: Do we change the sign of infinity if factor is negative?
> Again if we model this after the IEEE float behavior, then the answer
> is yes, we do change the sign of infinity.
>
> - Joe Koshakow
I ended up doing some more work in the attached patch. Here are some
updates:

- I modified the arithmetic operators to more closely match IEEE
floats. Error messages are still all TODO, and they may have the wrong
error code.
- I implemented some more operators and functions.
- I moved the helper functions you created into macros in timestamp.h
to more closely match the implementation of infinite timestamps and
dates. Also so dates.c could access them.
- There seems to be an existing overflow error with interval
subtraction. Many of the arithmetic operators of the form
`X - Interval` are converted to `X + (-Interval)`. This will overflow
in the case that some interval field is INT32_MIN or INT64_MIN.
Additionally, negating a positive infinity interval won't result in a
negative infinity interval and vice versa. We'll have to come up with
an efficient solution for this.

- Joe Koshakow