On 2017-11-19 19:49:01 -0500, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2017-11-19 19:08:48 -0500, Simon Riggs wrote:
> >> Am investigating the few buildfarm failures
>
> > The tests look very sensitive to differences in tuple size due to
> > different alignment requirements. Dependant on what MAXALIGN (and some
> > others) is the number of tuples fitting on a page will differ.
>
> "Few" buildfarm failures? It's probably going to fail on every 32-bit host.
>
> TBH, I would just remove those test cases. Even if they were stable
> across platforms, they don't directly prove anything at all about
> whether the feature does what it's supposed to.
>
> (Also, scaling the results to blocksize seems unlikely to help in
> passing on different BLCKSZ configurations...)
I think it might make sense to rewrite the tests so it doesn't output
any of the sizes, but instead just compares the size of tables with
different thresholds. That should be fairly reliable.
Greetings,
Andres Freund