On Thu, Mar 2, 2017 at 11:48 AM, Andres Freund <andres@anarazel.de> wrote:
> On 2017-03-01 19:25:23 -0600, Jim Nasby wrote:
>> On 2/28/17 11:21 AM, Andreas Karlsson wrote:
>> > The only downside I can see to this approach is that we no logner will
>> > able to reindex catalog tables concurrently, but in return it should be
>> > easier to confirm that this approach can be made work.
>>
>> Another downside is any stored regclass fields will become invalid.
>> Admittedly that's a pretty unusual use case, but it'd be nice if there was
>> at least a way to let users fix things during the rename phase (perhaps via
>> an event trigger).
>
> I'm fairly confident that we don't want to invoke event triggers inside
> the CIC code... I'm also fairly confident that between index oids
> stored somewhere being invalidated - what'd be a realistic use case of
> that - and not having reindex concurrently, just about everyone will
> choose the former.
Maybe. But it looks to me like this patch is going to have
considerably more than its share of user-visible warts, and I'm not
very excited about that. I feel like what we ought to be doing is
keeping the index OID the same and changing the relfilenode to point
to a newly-created one, and I attribute our failure to make that
design work thus far to insufficiently aggressive hacking.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
