Re: [HACKERS] Proposal: Local indexes for partitioned table

On Thu, Dec 7, 2017 at 4:57 PM, David Rowley
<david(dot)rowley(at)2ndquadrant(dot)com> wrote:
>> Sure, that would fix the problem I'm concerned about, but creating the
>> parent index first, as a shell, and then creating each child and
>> attaching it to the parent *also* fixes the problem I'm concerned
>> about, and without dragging any bystander objects into the fray.
>
> That's true, but is it worth inventing/maintaining an ATTACH syntax
> for that? It's not a very common case that multiple matching indexes
> existing. If it is worth it, do we need DETACH at all?

I think it is totally worth it. A little piece of extra DDL syntax
isn't that really costing us anything. Your proposed approach
probably still has obscure failure cases - e.g. I bet the
deadlock-avoidance logic in parallel restore won't know about the
dependency on a seemingly-unrelated object. Also, the use of a
seemingly-useless REPLACE syntax in every dump file will probably
confuse at least a few users, and maybe a few developers, too. I
think there is considerable value, both for the system and for the
humans who use it, in having something that has *exactly* the
semantics we want rather than *almost* the semantics we want.

I suppose if we want to get cute, we could have ONLY the ATTACH
syntax; if you attach an index for a partition that already has an
index attached, that could mean attach to the new one instead of the
old one (i.e. replace). But I would just add support for both ATTACH
and REPLACE and call it good.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company