Skip site navigation (1) Skip section navigation (2)

Re: Partitioning feature ...

From: Kedar Potdar <kedar(dot)potdar(at)gmail(dot)com>
To: Emmanuel Cecchet <manu(at)asterdata(dot)com>
Cc: pgsql-hackers(at)postgresql(dot)org, Amit Gupta <amit(dot)pc(dot)gupta(at)gmail(dot)com>
Subject: Re: Partitioning feature ...
Date: 2009-03-30 11:51:21
Message-ID: (view raw, whole thread or download thread mbox)
Lists: pgsql-hackers
Hi Emmanuel,

Thanks for your time. This is a WIP patch and we will integrate your
suggestions/comments as appropriate.


On Fri, Mar 27, 2009 at 3:38 AM, Emmanuel Cecchet <manu(at)asterdata(dot)com>wrote:

> Hi Kedar,
> First of all, congratulations for the excellent work.
> I have some comments and questions.
> In get_relevent_partition (btw, relevant is spelled with an a) you are
> maintaining 2 lists.

> Oops! 'a' typographical error.

I guess this is only useful for multi-column partitions, right?
> If you have a single column partition (without subpartitions), I think you
> could directly return on the first match (without maintaining any list)
> since you guarantee that there is no overlap between partitions.
> A simple but effective optimization for inserts consists of caching the
> last partition used (consecutive inserts often go to the same partition) and
> try it first before going through the whole loop.

> Yep.

> The update trigger should first check if the tuple needs to be moved. If
> the updated tuple still matches the constraints of the partitions it will
> not have to be moved and will save a lot of overhead.

> Yes. We agree on that.

> The COPY operation should probably be optimized to use the same code as the
> one in the insert trigger for partitioned tables. I guess some code could be
> factorized in COPY to make the inserts more efficient.
> The current trigger approach should prevent other triggers to be added to
> the table, or you should make sure that the partition trigger is always the
> one to execute last.

>  As triggers are executed in order of their names, we've prefixed the
trigger names with "zz". This should work fine as long as no-one uses
trigger-name which starts with "zz".

> As we don't have automatic partition creation, it would be interesting to
> have an optional mechanism to deal with tuples that don't match any
> partition (very useful when you do bulk insert and some new data require a
> new partition). Having a simple overflow partition or an error logging
> mechanism would definitely help to identify these tuples and prevent things
> like large COPY operations to fail.

> Will get back on this.

> Looking forward to your responses,
> Emmanuel
>> We are implementing table partitioning feature to support Range and Hash
>> partitions. Please find attached, the WIP patch and test-cases.
>> The syntax used conforms to most of the suggestions mentioned in
>> barring the following:
>> -- Specification of partition names is optional. System will be able to
>> generate partition names in such cases.
>> -- Sub partitioning
>> We are maintaining a system catalog(pg_partition) for partition meta-data.
>> System will look-up this table to find appropriate partition to operate on.
>> System internally uses low-level 'C' triggers to row-movement.
>> Regards,
>> --
>> Kedar.
>> ------------------------------------------------------------------------
> --
> Emmanuel Cecchet
> Aster Data Systems
> Web:

In response to


pgsql-hackers by date

Next:From: Merlin MoncureDate: 2009-03-30 11:55:24
Subject: Re: PQinitSSL broken in some use casesf
Previous:From: Marko KreenDate: 2009-03-30 10:11:48
Subject: 8.3.5: Crash in CountActiveBackends() - lockless race?

Privacy Policy | About PostgreSQL
Copyright © 1996-2017 The PostgreSQL Global Development Group