Hi,
On Thu, Mar 31, 2022 at 06:08:01PM +0300, Anton A. Melnikov wrote:
> Hello!
>
> On 30.03.2022 22:36, Robert Haas wrote:
> > I don't think that the idea of "extra" WAL records is very principled.
> > It's pretty vague what "extra" means, and your definition seems to be
> > basically "whatever would be needed to make this test case pass."
I agree, and even it there was a better definition there probably isn't much to
learn from it.
> I
> > think the problem is basically with the test cases's idea that # of
> > WAL records and # of table rows ought to be equal. I think that's just
> > false. In general, we'd also have to worry about index insertions,
> > which would provoke variable numbers of WAL records depending on
> > whether they cause a page split. And we'd have to worry about TOAST
> > table insertions, which could produce different numbers of records
> > depending on the size of the data, the configured block size and TOAST
> > threshold, and whether the TOAST table index incurs a page split.
Indeed, we added this test as it was hitting only a few queries with small
rows, which we thought would be stable, but that's apparently not the case. I
think the reason we never had any problem is that the buildfarm currently
doesn't run pg_stat_statement regression test, as it's marked as
NO_INSTALLCHECK. Other CI systems like at pgpro evidently have a different
approach.
> > If it's true that this test case sometimes randomly fails, then we
> > ought to fix that somehow, maybe by just removing this particular
> > check from the test case, or changing it to >=, or something like
> > that. But I don't think adding a new counter is the right idea.
>
> Indeed. Then there is a very simple solution for this particular case as
> wal_records counter may only sometime becomes greater but never less.
> The corresponding patch is attached.
+1 for this approach, and the patch looks good to me.
