Hello,
07.06.2024 17:25, Tom Lane wrote:
> Andrew Dunstan <andrew@dunslane.net> writes:
>> I still think my patch to force TCP mode for the SSL test makes sense as
>> well.
> +1 to both things. If that doesn't get the failure rate down to an
> acceptable level, we can look at the retry idea.
I'd like to add that the kerberos/001_auth test also suffers from the port
conflict, but slightly differently. Look for example at [1]:
krb5kdc.log contains:
Jul 02 09:29:41 andres-postgres-buildfarm-v5 krb5kdc[471964](info): setting up network...
Jul 02 09:29:41 andres-postgres-buildfarm-v5 krb5kdc[471964](Error): Address already in use - Cannot bind server socket
on 127.0.0.1.55853
Jul 02 09:29:41 andres-postgres-buildfarm-v5 krb5kdc[471964](Error): Failed setting up a UDP socket (for
127.0.0.1.55853)
Jul 02 09:29:41 andres-postgres-buildfarm-v5 krb5kdc[471964](Error): Address already in use - Error setting up network
As far as I can see, the port for kdc is chosen by
PostgreSQL::Test::Kerberos, via
PostgreSQL::Test::Cluster::get_free_port(), which checks only for TCP
port availability (with can_bind()), but not for UDP, so this increases
the probability of the conflict for this test (a similar failure: [2]).
Although we can also find a failure with TCP: [3]
(It's not clear to me, what processes can use UDP ports while testing,
but maybe those buildfarm animals are running on the same logical
machine simultaneously?)
[1] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=rorqual&dt=2024-07-02%2009%3A27%3A15
[2] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=mylodon&dt=2024-05-15%2001%3A25%3A07
[3] https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=grassquit&dt=2024-07-04%2008%3A28%3A19
Best regards,
Alexander
