Re: pg_dump crash on identity sequence with not loaded attributes

Artur Zakirov <zaartur(at)gmail(dot)com> writes:
> I think it will still be necessary to negate DUMP_COMPONENT_DEFINITION
> from seqinfo->dobj.dump because we shouldn't dump statements like ...
> ALTER COLUMN ... ADD GENERATED ..., if the table's definition isn't
> dumped.

Yeah. After chewing on this for awhile, I think the cleanest solution
is say that b965f2617 was just wrong, and we should revert it in
favor of adopting this logic:

if (seqinfo->is_identity_sequence)
seqinfo->dobj.dump = owning_tab->dobj.dump;
else
seqinfo->dobj.dump |= owning_tab->dobj.dump;

That is, for an identity sequence we should dump the exact same
properties as for its owning table. This basically rejects the
notion that a table can be inside an extension while its identity
sequence is not (or vice versa), which is the same position we
take for indexes, rules, etc. However, for non-identity owned
sequences (i.e. serials) keep the same behavior as before.
The large comment block just below here (added by f9e439b1c)
carefully describes the behavior that was wanted for non-identity
sequences, and I don't think we need to break that.

The attached patch also gets rid of the dubious coding in
getPublicationNamespaces. We might get push-back on that ignoring
schemas belonging to extensions, but if so I'd prefer to see the
behavior coded in a more transparent fashion.

regards, tom lane