Re: Proposal: Local indexes for partitioned table

On 2017/10/24 1:15, Alvaro Herrera wrote:
> Robert Haas wrote:
>> On Mon, Oct 23, 2017 at 11:12 AM, Alvaro Herrera
>> <alvherre(at)alvh(dot)no-ip(dot)org> wrote:
>>> I started with Maksim's submitted code, and developed according to the
>>> ideas discussed in this thread. Attached is a very WIP patch series for
>>> this feature.

Nice!

>>> Many things remain to be done before this is committable: pg_dump
>>> support needs to be written. ALTER INDEX ATTACH/DETACH not yet
>>> implemented. No REINDEX support yet. Docs not updated (but see the
>>> regression test as a guide for how this is supposed to work; see patch
>>> 0005). CREATE INDEX CONCURRENTLY not done yet.
>>>
>>> I'm now working on the ability to build unique indexes (and unique
>>> constraints) on top of this.
>>
>> Cool. Are you planning to do that by (a) only allowing the special
>> case where the partition key columns/expressions are included in the
>> indexed columns/expressions, (b) trying to make every insert to any
>> index check all of the indexes for uniqueness conflicts, or (c)
>> implementing global indexes? Because (b) sounds complex - think about
>> attach operations, for example - and (c) sounds super-hard. I'd
>> suggest doing (a) first, just on the basis of complexity.
>
> Yes, I think (a) is a valuable thing to have -- not planning on doing
> (c) at all because I fear it'll be a huge time sink. I'm not sure about
> (b), but it's not currently on my plan.

+1 to proceeding with (a) first.

>> I hope that you don't get so involved in making this unique index
>> stuff work that we don't get the cascading index feature, at least,
>> committed to v11. That's already a considerable step forward in terms
>> of ease of use, and I'd really like to have it.
>
> Absolutely -- I do plan to get this one finished regardless of unique
> indexes.

+1

Thanks,
Amit