On Tue, Dec 18, 2018 at 1:20 PM Alvaro Herrera <alvherre@2ndquadrant.com> wrote:
>
> On 2018-Dec-18, Peter Geoghegan wrote:
> Hmm, interesting. I wonder if this is just a case of never testing this
> code under "postgres --ignore-system-indexes".
I suppose that you could say that. The regression tests will fail at
many points with --ignore-system-indexes, almost all of which are due
to well understood issues. For example, you'll get load of WARNINGs
about needing to use a system index despite the server being run under
--ignore-system-indexes.
> I can reproduce the
> reported problem without your patch by using that flag. Here's a
> recipe:
>
> create extension cube;
> create table dep as select ctid as tid,* from pg_depend;
> create extension earthdistance;
> select tid, deptype, (dep).type, (dep).identity, (ref).type, (ref).identity
> from (select tid, deptype, pg_identify_object(classid, objid, objsubid) as dep,
> pg_identify_object(refclassid, refobjid, refobjsubid) as ref
> from (select ctid as tid, * from pg_depend except select * from dep) a
> ) b;
Interesting.
Note that if the standard that we're going to hold a solution to here
is "must produce sane output with --ignore-system-indexes", then my
solution will not meet that standard. However, I fear that it's going
to be really hard to accomplish that goal some other way (besides
which, as I said, the tests will still fail with
--ignore-system-indexes for reasons that have nothing to do with
dependency management).
--
Peter Geoghegan