Re: [bug?] Missed parallel safety checks, and wrong parallel safety

On Tue, May 4, 2021 at 11:47 PM Greg Nancarrow <gregn4422(at)gmail(dot)com> wrote:
> Problem is, for built-in functions, the changes are allowed, but for
> some properties (like strict) the allowed changes don't actually take
> effect (this is what Amit was referring to - so why allow those
> changes?).
> It's because some of the function properties are cached in
> FmgrBuiltins[] (for a "fast-path" lookup for built-ins), according to
> their state at build time (from pg_proc.dat), but ALTER FUNCTION is
> just changing it in the system catalogs. Also, with sufficient
> privileges, a built-in function can be redefined, yet the original
> function (whose info is cached in FmgrBuiltins[]) is always invoked,
> not the newly-defined version.

I agree. I think that's not ideal. I think we should consider putting
some more restrictions on updating system catalog changes, and I also
think that if we can get out of having strict need to be part of
FmgrBuiltins[] that would be good. But what I don't agree with is the
idea that since strict already has this problem, it's OK to do the
same thing with parallel-safety. That seems to me to be making a bad
situation worse, and I can't see what problem it actually solves.

--
Robert Haas
EDB: http://www.enterprisedb.com