Horton

Re: Multiple hosts in connection string failed to failover in non-hot standby mode

Tom Lane <tgl@sss.pgh.pa.us>

View thread View on postgr.es
From: Tom Lane <tgl@sss.pgh.pa.us>
To: Michael Paquier <michael@paquier.xyz>
Cc: Justin Pryzby <pryzby@telsasoft.com>, Hubert Zhang <zhubert@vmware.com>, tsunakawa.takay@fujitsu.com, pgsql-hackers@postgresql.org, Andreas Seltenreich <seltenreich@gmx.de>, Bruce Momjian <bruce@momjian.us>
Date: 2021-05-31T13:36:20Z
Lists: pgsql-hackers 
Michael Paquier <michael@paquier.xyz> writes:
> On Mon, May 31, 2021 at 12:05:12AM -0400, Tom Lane wrote:
>> What is not clear is why GSS is acting that way.  We wouldn't
>> have tried a GSS connection unless pg_GSS_have_cred_cache
>> succeeded ... so how come that worked but then gss_init_sec_context
>> complained "Credential cache is empty"?

> I suspect that's just the way the upstream installation works with a
> credentials cache created from the beginning, as I see the same
> behavior and the same error on my own host for HEAD with a KRB5 server
> set up once the upstream installation runs.

Interesting --- I was considering running such a test locally, but
didn't get round to it yet.

> Leaving the specific
> topic of this thread aside for one moment, would there be an argument
> for just enforcing gssencmode=disable in this set of tests down to 12?

It seems like the ideal solution would be to make pg_GSS_have_cred_cache
smarter, so that we don't attempt a GSS connection cycle here.  But if
we can't, adding gssencmode=disable to these test cases is what I was
thinking about, too.

> Another thing that strikes me as incorrect is that we don't unset
> PGGSSENCMODE or PGGSSLIB in TestLib.pm.  Just noting it on the way..

Agreed, that seems bogus.

			regards, tom lane

Commits

  1. Clear conn->errorMessage at successful completion of PQconnectdb().

  2. Avoid ECPG test failures in some GSS-capable environments.

  3. Try next host after a "cannot connect now" failure.

  4. Uniformly identify the target host in libpq connection failure reports.

  5. Allow pg_regress.c wrappers to postprocess test result files.

  6. In libpq, always append new error messages to conn->errorMessage.