On Mon, 2011-06-06 at 14:42 +0200, Dimitri Fontaine wrote:
> I think the way things are going to be organised now is that we will
> have core-blessed extensions: don't mix the mechanism and the policy.
I like that idea.
> > non-issue if we had a good type interface system (that works on
> > polymorphic types) -- we could just have a built-in "range" interface,
> > and the range extension could add "&&" as the range interface's overlaps
> > operator for the type ANYRANGE.
>
> That means that this is, IMHO, the right approach. Have core support
> that enables user defined RANGE types with indexing and planner support,
> etc, like we have OPERATOR CLASS and FAMILY and all the jazz.
If we take the minimal approach, the index support would be the first to
be moved to an extension. In order to have index support in core, we
need quite a few functions and a significant amount of code.
Minimal would be: * CREATE TYPE ... AS RANGE * ANYRANGE * The IO functions * Possibly the constructors and accessors (
range(), range_oc(), lower(), upper(), etc.)
Regarding the type interfaces, the only thing that really worries me
there is that my future work will depend on them existing, and I haven't
really thought through the details. For instance, it just occurred to me
recently that it would need to support polymorphic types, which might be
a little bit more complicated than a simple lookup.
I suppose it's easier to put a few functions in core later if we get
stuck than to rip them out later.
> And the useful stuff you need to have to benefit from that core support
> would be an extension. It could be a core maintained extension, and it
> could even get installed by default, so that all the users would need to
> do is 'CREATE EXTENSION timeranges;', for example.
Sounds good to me. However, would the extension be available in
pg_regress? If not, I will need to include those constructors/accessors
to be able to test anything.
> I think the consensus is to instead add a new chapter (maybe between
> current chapters 9. Functions and Operators and 10. Type Conversion) and
> host “core extensions” docs there. The source code organisation is
> controversial because technically not necessary. We have to keep the
> work Greg did to keep those contribs shipped by default. Oh, and that
> is on the 9.1 Open Items, right?
OK, so there are still a few things to be decided around documentation
and tests. Both of those things can take a significant amount of time to
rework, so I think I'll leave it alone until we have more of a
consensus.
We still have time before 9.2 to break some of the code out into an
extension when we do have the doc/test issues resolved.
Regards,Jeff Davis
