Re: [HACKERS] Drop one-argument string_agg? (was Re: string_agg delimiter having no effect with order by)

I wrote:
> Hm? I don't think that an initdb here would have any impact on whether
> we can call the next drop RC1 or not. We're talking about removing a
> single built-in entry in pg_proc --- it's one of the safest changes we
> could possibly make.

Well, I forgot that an aggregate involves more than one catalog row ;-).
So it's a bit bigger patch than that, but still pretty small and safe.
See attached.

What we are doing here, IMO, is not just changing string_agg() but
instituting a project policy that we are not going to offer built-in
aggregates with the same names and different numbers of arguments ---
otherwise the problem will come right back. Therefore, the attached
patch adds an opr_sanity regression test that will complain if anyone
tries to add such. Of course, this isn't preventing users from creating
such aggregates, but it's on their own heads to not get confused if they
do.

This policy also implies that we are never going to allow default
arguments for aggregates, or at least never have any built-in ones
that use such a feature.

By my count the following people had offered an opinion on making
this change:
for: tgl, josh, badalex, mmoncure
against: rhaas, thom
Anybody else want to vote, or change their vote after seeing the patch?

regards, tom lane