I wouldn’t discuss the algorithm on this list. I’d just commit to being compatible with Postgres, and write a bunch of tests based on Postgres’ observed behavior.
> On Sep 23, 2021, at 5:12 AM, Phillip Cloud <cpcl...@gmail.com> wrote: > > Hi all, > > I wanted to draw some attention to ARROW-11090 [1] in an effort to start > getting the implementation going for temporal arithmetic kernels. > > The supported operation matrix is decently large. > > I would like to propose that we adopt the behavior of PostgreSQL where > things are at first glance ambiguous, such as this operation: > > SELECT DATE '2021-05-30' + INTERVAL '1' MONTH > > I couldn't find any documentation about this specific case, but the code is > fairly straightforward and has a useful comment describing the behavior > [2]. The timestamp_pl_interval will first add 1 to the month number [3], > and then take the minimum of the month day (in this case 31, for the > non-existent date 2021-06-31 and the latest valid day in the month (the > latest valid day being 30; month day counts are encoded as a table with two > rows one for leap years one for non leap years [4]) [5]. > > Thoughts? > > [1]: https://issues.apache.org/jira/browse/ARROW-11090 > [2]: > https://github.com/postgres/postgres/blob/master/src/backend/utils/adt/timestamp.c#L2830-L2839 > [3]: > https://github.com/postgres/postgres/blob/master/src/backend/utils/adt/timestamp.c#L2862 > [4]: > https://github.com/postgres/postgres/blob/master/src/interfaces/ecpg/pgtypeslib/dt_common.c#L14-L16 > [5]: > https://github.com/postgres/postgres/blob/master/src/backend/utils/adt/timestamp.c#L2874-L2876
