Re: BUG #16643: PG13 - Logical replication - initial startup never finishes and gets stuck in startup loop

Alvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org> writes:
> On 2020-Oct-01, Peter Eisentraut wrote:
>> What's the difference between this case and what the test suite is testing?
>> Is it that it replicates between two databases on the same instance?

> I don't know why the tests pass, but the message
> ERROR: error reading result of streaming command:
> does appear in the logs after running src/test/subscription many times
> (I see it in tests 001, 002, 013 and 014, apart from the new one in
> 100). It's indeed surprising that these tests all pass!

> I turned Henry's reproducer into the attached TAP test, and it does
> reproduce the problem; but if I reduce the number of rows from 5000 to
> 1000, then it no longer does. I don't quite see why this would be a
> problem with a larger table only. Do you?

I think we really need to figure that out before concluding that this
problem is solved. Now that we've seen this, I'm wondering urgently
what other coverage gaps we've got there.

> The fix is the commented-out line in walsender.c; the test reliably
> passes for me if I uncomment that, and the error message disappear from
> the server logs in all the other tests.

I agree that this is what we need to do code-wise; we can't let the
protocol break stand, or we'll break all sorts of cross-version
replication scenarios. However, we'd better also change the protocol
spec to say that this is what is supposed to happen.

regards, tom lane