Thread: regression tests fail for different block sizes

regression tests fail for different block sizes

From
Jeff Davis
Date:
In master, 4 tests fail due to plan changes with blocksize 32K. The
failures started creeping in around 9.0.

I don't see a clean way to make the plans consistent with 8K and 32K
pages, so I was about to write this off as "we don't care much".

Then I ran with block size 1KB, and I got what appears to be a genuine
failure related to range types and/or spgist. I'm still investigating
that, which will be a separate discussion, but the fact that different
block sizes can catch real errors makes me think we should properly
support them in the regression tests.

Any ideas how we can make the plans consistent enough when the blocksize
varies?

Regards,Jeff Davis





Re: regression tests fail for different block sizes

From
Tom Lane
Date:
Jeff Davis <pgsql@j-davis.com> writes:
> In master, 4 tests fail due to plan changes with blocksize 32K. The
> failures started creeping in around 9.0.

AFAIR, that's been true since the dark ages, not since 9.0.

> I don't see a clean way to make the plans consistent with 8K and 32K
> pages, so I was about to write this off as "we don't care much".

Indeed.  There are any number of parameters that you can change that will
break the regression tests by changing plan choices.  If we tried to
cover all such cases we'd end up with lobotomized tests that detect
fewer issues, not more.

> Then I ran with block size 1KB, and I got what appears to be a genuine
> failure related to range types and/or spgist.

This might be worth investigating, but it doesn't mean that we need
automated regression tests covering the case.
        regards, tom lane