Re: tests against running server occasionally fail, postgres_fdw & tenk1

Hi,

On 2023-02-06 19:29:46 -0800, Andres Freund wrote:
> There's something off. Isolationtester's control connection emits *loads* of
> invalidation messages:
> 2023-02-06 19:29:06.430 PST [2125297][client backend][6/0:121864][isolation/receipt-report/control connection] LOG: previously emitted 7662 messages, 21 this time
> 2023-02-06 19:29:06.566 PST [2125297][client backend][6/0:121873][isolation/receipt-report/control connection] LOG: previously emitted 8155 messages, 99 this time
> 2023-02-06 19:29:06.655 PST [2125297][client backend][6/0:121881][isolation/receipt-report/control connection] LOG: previously emitted 8621 messages, 99 this time
> 2023-02-06 19:29:06.772 PST [2125297][client backend][6/0:121892][isolation/receipt-report/control connection] LOG: previously emitted 9208 messages, 85 this time
> 2023-02-06 19:29:06.867 PST [2125297][client backend][6/0:121900][isolation/receipt-report/control connection] LOG: previously emitted 9674 messages, 85 this time

Ah, that's just due to setup-teardown happening in the control connection.

FWIW, I see plenty of sinval resets even if I increase the sinval queue size
substantially. I suspect we've increased the number of sinval messages we sent
a good bit over time, due to additional syscaches.

As we only process catchup interrupts while in ReadCommand(), it's not
surprising that we very quickly get behind. A single statement suffices.

Anyway, that's not really a correctness issue, just a performance issue.

But the postgres_fdw.sql vulnerability to syscache resets seems not
great. Perhaps pgfdw_inval_callback() could check if the definition of the
foreign server actually changed, before dropping the connection?

Greetings,

Andres Freund