Re: SIGSEGV from START_REPLICATION 0/XXXXXXX in XLogSendPhysical () at walsender.c:2762

Hi,

On 2020-06-03 18:27:12 -0400, Alvaro Herrera wrote:
> On 2020-Jun-03, Andres Freund wrote:
> > I don't think we should prohibit this. For one, it'd probably break some
> > clients, without a meaningful need.
>
> There *is* a need, namely to keep complexity down. This is quite
> convoluted, it's got a lot of historical baggage because of the way it
> was implemented, and it's very difficult to understand. The greatest
> motive I see is to make this easier to understand, so that it is easier
> to modify and improve in the future.

That seems like a possibly convincing argument for not introducing the
capability, but doesn't seem strong enough to remove it. Especially not
if it was just broken as part of effectively a refactoring, as far as I
understand?

> > But I think it's also actually quite useful to be able to access
> > catalogs before streaming data. You e.g. can look up configuration of
> > the primary before streaming WAL. With a second connection that's
> > actually harder to do reliably in some cases, because you need to be
> > sure that you actually reached the right server (consider a pooler,
> > automatic failover etc).
>
> I don't think having a physical replication connection access catalog
> data directly is a great idea. We already have gadgets like
> IDENTIFY_SYSTEM for physical replication that can do that, and if you
> need particular settings you can use SHOW (commit d1ecd539477). If
> there was a strong need for even more than that, we can add something to
> the grammar.

Those special case things are a bad idea, and we shouldn't introduce
more. It's unrealistic that we can ever make that support everything,
and since we already have to support the database connected thing, I
don't see the point.

Greetings,

Andres Freund