On Thu, Jul 23, 2020 at 01:05:04PM -0400, Jeff Janes wrote:
> I have noticed this before, but since it wasn't a production machine I just
> shrugged it off as being a hazard of using consumer-grade stuff; it didn't
> seem to be worth investigating further.
The most direct and non-invasive way to address this problem in
back-branches would be to copy the non-concurrent logic in
src/port/open.c directly into pg_test_fsync.c. We have done that in
the past such things for example with restricted tokens on Windows for
pg_upgrade, see fa1e5afa but that was much more important. I am not
sure that if it is worth the time spent here though, as it took
roughly 10 years to notice the problem (and I don't really count the
time where the tool was under src/tools/ but this did not build the
tool on Windows).
--
Michael