Re: Extract epoch from Interval weird behavior

On 24.02.22 03:35, Joseph Koshakow wrote:
>> However when executing EXTRACT we first truncate
>> DAYS_PER_YEAR to an integer, and then multiply it
>> by the total years in the Interval
>> /* this always fits into int64 */
>>> secs_from_day_month = ((int64) DAYS_PER_YEAR * (interval->month / MONTHS_PER_YEAR) +
>>> (int64) DAYS_PER_MONTH * (interval->month % MONTHS_PER_YEAR) +
>>> interval->day) * SECS_PER_DAY;
>> Is this truncation on purpose? It seems like
>> EXTRACT is not accounting for leap years in
>> it's calculation.

This was not intentional. The cast is only to make the multiplication
happen in int64; it didn't mean to drop any fractional parts.

> Oops I sent that to the wrong email. If this isn't intented I've created a patch
> that fixes it, with the following two open questions
> * DAYS_PER_YEAR_NUM is recalculated every time. Is there anyway
> to convert a float directly to a numeric to avoid this?

We really wanted to avoid doing calculations in numeric as much as
possible. So we should figure out a different way to write this. The
attached patch works for me. It's a bit ugly since it hardcodes some
factors. Maybe we can rephrase it a bit more elegantly.