Hi,
On Thu, Apr 8, 2021, at 07:52, Tom Lane wrote:
> Andrew Dunstan <andrew@dunslane.net> writes:
> > On 4/7/21 9:50 PM, Andres Freund wrote:
> >> I've wondered about that too. Perhaps we could reuse the pg_upgrade run?
>
> > Honestly I'd prefer it if we could get rid of the rerun of 'make check'
> > by pg_upgrade's test.sh and instead upgrade the data directory made by
> > the earlier 'make check' run if it's still there (which would mean we'd
> > need to stop it being deleted).
>
> Good idea as far as speeding check-world and buildfarm runs, but I wonder
> if we wouldn't be losing test coverage. Seeing the number of times that
> buildfarm runs have gotten through "make check" only to fail at the
> re-run in pg_upgrade, it seems clear to me that there is something
> different about the execution environment in the latter case. I've
> never been able to pin down quite what :-(
IIRC we made it run with fsync explicitly enabled. Which obviously will change timing somewhat substantially...
Andres