Greg Sabino Mullane <htamfids@gmail.com> writes:
> I'd echo the suggestion to strace this. You can use the pre_auth_delay
> setting to help facilitate that. See:
IIUC, the delays are rare and unpredictable, so that strace'ing seems
unlikely to be practical.
If rebuilding from source is feasible, you could insert monitoring
elog(LOG) calls around the name lookup (that is, around the
pg_getnameinfo_all call in src/backend/libpq/hba.c). This would
confirm or disprove the theory that the name lookup is the source
of the issue, which is about as much as we'd get out of strace.
regards, tom lane