On Tue, Sep 29, 2020 at 2:01 AM Tom Lane <t...@sss.pgh.pa.us> wrote:
> Amit Langote <amitlangot...@gmail.com> writes:
> > On Fri, Sep 25, 2020 at 12:02 AM Tom Lane <t...@sss.pgh.pa.us> wrote:
> >> Well, I agree with Peter to that extent, but my opinion is that *none*
> >> of these cases ought to be errors. What we're doing here is performing
> >> an implicit assignment-level coercion of the expression to the type of
> >> the column, and changing the collation is allowed as part of that:
> >>
> >> regression=# create table foo (f1 text collate "C");
> >> CREATE TABLE
> >> regression=# insert into foo values ('a' COLLATE "POSIX");
> >> INSERT 0 1
> >> regression=# update foo set f1 = 'b' COLLATE "POSIX";
> >> UPDATE 1
> >>
> >> So I find it completely inconsistent that the partitioning logic
> >> complains about equivalent cases.
>
> > My perhaps wrong impression was that the bound expression that is
> > specified when creating a partition is not as such being *assigned* to
> > the key column, but now that I think about it some more, that doesn't
> > matter.
>
> >> I think we should just rip the
> >> whole thing out, as per the attached draft. This causes several
> >> regression test results to change, but AFAICS those are only there
> >> to exercise the error tests that I think we should get rid of.
>
> > Yeah, I can see no other misbehavior resulting from this.
>
> OK, I'll clean up the regression test cases and push that.
Thanks.
> (Although this could be claimed to be a bug, I do not feel
> a need to back-patch the behavioral change.)
Agreed. The assign_expr_collations() omission was indeed a bug.
--
Amit Langote
EnterpriseDB: http://www.enterprisedb.com
