Thomas Munro <thomas.mu...@gmail.com> writes: > cfbot found another source of nondeterminism in the regression tests, > due to the smaller shared_buffers used in this TAP test:
This failure seems related but not identical: https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=myna&dt=2022-04-02%2004%3A00%3A26 portals.out is expecting that the "foo25ns" cursor will read starting at the beginning of tenk1, but it's starting somewhere else, which presumably is a syncscan effect. I think the fundamental instability here is that this TAP test is setting shared_buffers small enough to allow the syncscan logic to kick in where it does not in normal testing. Maybe we should just disable syncscan in this test script? regards, tom lane
