Reply-To: jesper.pedersen@redhat.com
Subject: Re: [HACKERS] Proposal: Local indexes for partitioned table
To: Alvaro Herrera <alvherre@alvh.no-ip.org>
Cc: Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>,
 Robert Haas <robertmhaas@gmail.com>,
 David Rowley <david.rowley@2ndquadrant.com>,
 Simon Riggs <simon@2ndquadrant.com>,
 Pg Hackers <pgsql-hackers@postgresql.org>,
 Maksim Milyutin <milyutinma@gmail.com>
References: <20180108203638.6di3uo7r7vqfsvda@alvherre.pgsql>
From: Jesper Pedersen <jesper.pedersen@redhat.com>
Organization: Red Hat
Message-ID: <3cc142e1-82c7-0d47-681b-59d519606f4e@redhat.com>
Date: Mon, 8 Jan 2018 15:47:15 -0500
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101
 Thunderbird/52.5.0
MIME-Version: 1.0
In-Reply-To: <20180108203638.6di3uo7r7vqfsvda@alvherre.pgsql>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Precedence: bulk

Hi Alvaro,

On 01/08/2018 03:36 PM, Alvaro Herrera wrote:
> Jesper Pedersen wrote:
> 
>> Maybe a warning for existing indexes on the same column(s), but with a
>> different type, should be emitted during ATTACH, e.g.
> 
>> [detach one partition, replace index with a different one, attach
>> partition]
> 
>> Of course, this could also fall under index maintenance and no warning
>> emitted. Docs have "Each partition is first checked to determine whether an
>> equivalent index already exists," so it is covered.
> 
> Yeah, I'm of two minds about this also -- in the initial versions I had
> a code comment wondering exactly about having a hash index in a
> partition attached to a btree index on parent.
> 
> As another example, having a documents table with two partitions (one
> "long term archival" and other "documents currently being messed with")
> you could have a text search index which is GIN in the former and GiST
> in the latter.  There is a performance argument for doing it that way,
> so it's not merely academic.
> 
> Anyway, while I think attaching an index that doesn't match the
> properties of the index on parent can be a useful feature, on the other
> hand it could be surprising (you end up losing an index because it was
> matched during attach that you didn't think was going to be matched).
> One idea would be to have a way to specify at ATTACH time what to do
> about indexes when they don't match exactly, but I think the user
> interface would be pretty messy: exactly how different do you want to
> allow the indexes to be?  Is an index having one more column than the
> one in parent good enough?  I think the answer to this is mostly a
> judgement call, and I'd rather not spend my time debating those.
> 

Yeah, agreed - lets leave as is.

Migrating an index to another type would mean to drop the top-level 
definition anyway.

Best regards,
  Jesper