On Fri, Jun 19, 2020 at 10:28:56AM -0700, Peter Geoghegan wrote:
> On Fri, Jun 19, 2020 at 10:27 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> I'm a bit skeptical about changing anything here. There are quite
>> a large number of GUCs that can affect the regression results, and
>> it wouldn't be sane to try to force them all to fixed values. For
>> one thing, that'd be a PITA to maintain, and for another, it's not
>> infrequently useful to run the tests with nonstandard settings to
>> see what happens.
>
> +1
We cared about such plan stability that in the past FWIW, see for
example c588df9 as work_mem is a setting that people like to change.
Why should this be different? work_mem is a popular configuration
setting. Perhaps people will not complain about that being an issue
if running installcheck, we'll know with the time. Anyway, I am fine
to just change my default configuration if the conclusion is to not
touch that and let it be, but I find a bit annoying that switching
work_mem from 4MB to 1MB is enough to destabilize the tests. And this
worked just fine in past releases.
--
Michael