Re: Unsynchronized parallel dumps from 13.3 replica produced by pg_dump

Thanks Tom. It's a strange one for sure. Hopefully AWS support will shed
some light on it. I will clarify too that this is the regular RDS Postgres
version and not their other Aurora Postgres service. I suspect the Aurora
Postgres probably differs from the community version by quite a bit, but
I'm unsure how much their regular Postgres offering differs, if at all.

Thanks,
Chris

On Mon, Oct 18, 2021 at 8:05 PM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Chris Williams <cswilliams(at)gmail(dot)com> writes:
> > We have a script that runs a pg_dump off of an RDS PG13.3 replica several
> > times per day. We then load this dump using pg_restore into another
> > postgres RDS db in another AWS account, scrub some of the data, and then
> > take a snapshot of it.
>
> Hmm ... I'm fairly sure that RDS Postgres is not Postgres at this level
> of detail. The info I've been able to find about their replication
> mechanism talks about things like "eventually consistent reads", which
> is not something community Postgres deals in.
>
> In particular, what I'd expect from the community code is that a replica
> could see a sequence as being *ahead* of the value that you might expect
> from looking at related tables; but never behind. (Also, that statement
> is true regardless of whether you are doing parallel dump.) And
> non-sequence tables should always be consistent, period.
>
> So I'm suspicious that this is an RDS-specific effect, and thus that
> you should consult Amazon support first. If they say "no, it's Postgres
> all the way down", then we need to look closer.
>
> regards, tom lane
>