Re: GiST support for inet datatypes

2014-02-19 16:52, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>:
> Not at all, AFAICS. If it were okay to decide that some formerly-default
> opclass is no longer default, then having such a command would be better
> than manually manipulating pg_opclass.opcdefault --- but extension upgrade
> scripts could certainly do the latter if they had to. The problem here
> is whether it's upgrade-safe to make such a change at all; having
> syntactic sugar doesn't make it safer.

It is not hard to support != operator on the new operator class. That would
make it functionally compatible with the ones on btree_gist. That way,
changing the default would not break pg_dump dumps, it would only upgrade
the index to the new one.

pg_upgrade dumps current objects in the extension. It fails to restore them,
if the new operator class exists as the default. I do not really understand
how pg_upgrade handle extension upgrades. I do not have a solution to this.

It would be sad if we cannot make the new operator class default because of
the btree_gist implementation which is not useful for inet data types. You
wrote on 2010-10-11 about it [1]:

> Well, actually the btree_gist implementation for inet is a completely
> broken piece of junk: it thinks that convert_network_to_scalar is 100%
> trustworthy and can be used as a substitute for the real comparison
> functions, which isn't even approximately true. I'm not sure why
> Teodor implemented it like that instead of using the type's real
> comparison functions, but it's pretty much useless if you want the
> same sort order that the type itself defines.

[1] http://www.postgresql.org/message-id/8973.1286841006@sss.pgh.pa.us