Re: Fixing findDependentObjects()'s dependency on scan order (regressions in DROP diagnostic messages)

On Fri, Jan 18, 2019 at 4:06 PM Peter Geoghegan <pg(at)bowt(dot)ie> wrote:
> > * Objects-to-drop output from DROP ROLE is still unstable. I suppose
> > this would be fixed by also doing sorting in that code path, but
> > I've not looked into it.
>
> The nbtree patch is not dependent on doing better here, since the
> order here is merely unstable, without leading to incorrect diagnostic
> messages. I can just change the regress output mechanically. That
> said, I cannot be sure that the instability won't become more of an
> annoyance with heap TID being treated as a tie-breaker attribute
> within nbtree. It's probably fine to reserve the option to do better
> here, and do so if and when it becomes clear that it matters. I defer
> to you.

Good news: Testing my patch (with the hacky pg_depend tie-breaker
commit removed) on top of your recent commit f1ad067f confirms the
only "real" remaining problem is with two tests that relate to
partitioning (the multiple DEPENDENCY_INTERNAL_AUTO entry issue). It
seems likely that objects-to-drop output from DROP ROLE can remain
"unstable" without bothering anybody -- I've run "make -Otarget -j10
-s check-world" with the same source tree multiple times, and have yet
to see a test failure. In practice *all* of the instability that's of
practical concern (that could cause buildfarm failures) related to the
two pg_depend indexes. It very much looks that way, at least -- only
the buildfarm can confirm this.

I attach a patch that shows how I'll adjust the harmless
objects-to-drop output from DROP ROLE to make relevant tests pass --
the scope of changes is very limited, and the changes are
harmless/mechanical. Separately, I attach the patch that I'd have to
use to paper over the partitioning/DEPENDENCY_INTERNAL_AUTO issue (a
patch that makes the regression tests actively ignore visible
manifestations of the DEPENDENCY_INTERNAL_AUTO bug). As I said, this
confirms that there are only two "real" remaining tests that fail.

Thanks
--
Peter Geoghegan