Re: NOTIFY does not work as expected

Hi,

On 2018-10-19 19:53:06 -0400, Tom Lane wrote:
> Andres Freund <andres(at)anarazel(dot)de> writes:
> > On 2018-10-19 13:45:42 -0700, Andres Freund wrote:
> >> I don't immediately see a problem with changing this for reads.
>
> > One argument against changing it, although not a very strong one, is
> > that processing a proc die even when non-blocking prevents us from
> > processing commands like a client's X/terminate even if we already have
> > the necessary input.
>
> I'm pretty skeptical of these arguments, as they depend on assumptions
> that there are no CHECK_FOR_INTERRUPTS calls anywhere in the relevant
> code paths outside be-secure.c. Even if that's true today, it doesn't
> seem like something to depend on.

I'm not particularly convinced either, if you're just talking about
reads, rather than writes. I think it'd matter more if we had a mode
where we told clients "dear clients please shut down now, I'm about to
do the same" and we wanted to avoid redundant log-spam about client EOF
etc, but we don't.

> However, there's definitely merit in the idea that we shouldn't change
> the ProcDie behavior if we don't have to in order to fix the NOTIFY
> bug --- especially since I'd like to backpatch this. So if you're
> happy with the revised patch, I can go with that one.

I find the +1, 0, -1 pretty confusing and less clear than what was there
before. If we do something like it, could we introduce an enum or macros
for it?

Greetings,

Andres Freund