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

Re: [HACKERS] Including Snapshot Info with Indexes

From: Bruce Momjian <bruce(at)momjian(dot)us>
To: Gokulakannan Somasundaram <gokul007(at)gmail(dot)com>
Cc: pgsql-patches(at)postgresql(dot)org, Luke Lonergan <llonergan(at)greenplum(dot)com>, Gregory Stark <stark(at)enterprisedb(dot)com>, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>, Andreas Joseph Krogh <andreak(at)officenet(dot)no>, Hannu Krosing <hannu(at)skype(dot)net>, pgsql-hackers(at)postgresql(dot)org
Subject: Re: [HACKERS] Including Snapshot Info with Indexes
Date: 2008-04-10 18:56:08
Message-ID: 200804101856.m3AIu8523356@momjian.us (view raw or flat)
Thread:
Lists: pgsql-hackerspgsql-patches
I have added URLs to your patch to the TODO list:

	* Allow data to be pulled directly from indexes

---------------------------------------------------------------------------

Gokulakannan Somasundaram wrote:
> Hi,
>     I would like to present the first patch. It currently has the following
> restrictions
> a) It does not support any functional indexes.
> b) It supports queries like select count(1) from table where (restrictions
> from indexed columns), but it does not support select count(1) from table.
> 
> The Syntax to create this type of index is
> 
> create thick index idx on dd(n1,n2)
> 
> here idx- index name and dd- table name and n1 and n2 are column names.
> 
> I have created a extra column in pg_index called indhassnapshot.
> 
> I have also enabled the display of Logical Reads. In order to see that, set
> log_statement_stats on.
> 
> The thick index is clearly on the front, if you issue queries like
> 
> select n2 from dd where n1>1000 and n2<1500;
> 
> As already said, if the update is not incurring any extra cost, except if
> the indexed columns are updated. Deletes are costly, making it ideal for
> partitioned tables.
> 
> In order to update the thick indexes, i have accessed the ps_ExprContext in
> PlanState to get the oldtuple. But if we have a outer plan and inner plan,
> then i have set the ps_ExprContext of  innerplan to the outerplan. I don't
> know whether there will be instances where the ps_ExprContext of outerplan
> node will have some use in update queries.
> 
> Right now, it passes the regression test suite. I had slight trouble with
> pg_indent, so i think it has not got applied properly. But i have tried to
> remove all the whitespace differences. Please be kind to me in case i have
> missed any whitespace differences. :)
> 
> Please review the patch and provide your comments.
> 
> Thanks,
> Gokul.
> CertoSQL Project,
> Allied Solution Groups.
> (www.alliedgroups.com)
> 
> On 10/23/07, Hannu Krosing <hannu(at)skype(dot)net> wrote:
> >
> > ?hel kenal p?eval, L, 2007-10-20 kell 10:19, kirjutas Luke Lonergan:
> > > Hi Hannu,
> > >
> > > On 10/14/07 12:58 AM, "Hannu Krosing" <hannu(at)skype(dot)net> wrote:
> > >
> > > > What has happened in reality, is that the speed difference between
> > CPU,
> > > > RAM and disk speeds has _increased_ tremendously
> > >
> > > Yes.
> > >
> > > > which makes it even
> > > > more important to _decrease_ the size of stored data if you want good
> > > > performance
> > >
> > > Or bring the cpu processing closer to the data it's using (or both).
> > >
> > > By default, the trend you mention first will continue in an unending way
> > -
> > > the consequence is that the "distance" between a processor and it's
> > target
> > > data will continue to increase ad-infinitum.
> >
> > the emergence of solid-state (flash) disks may help a little here, but
> > in general it is true.
> >
> > > By contrast, you can only decrease the data volume so much - so in the
> > end
> > > you'll be left with the same problem - the data needs to be closer to
> > the
> > > processing.  This is the essence of parallel / shared nothing
> > architecture.
> > >
> > > Note that we've done this at Greenplum.  We're also implementing a
> > DSM-like
> > > capability and are investigating a couple of different hybrid row /
> > column
> > > store approaches.
> >
> > Have you tried moving the whole visibility part of tuples out to a
> > separate heap ?
> >
> > Especially in OLAP/ETL scenarios the distribution of tuples loaded in
> > one transaction should be very good for visibility-info compression.
> >
> > I'd suspect that you could crush hundreds of pages worth of visibility
> > into single RLE encoding unit (xmin=N, xmax=no_yet, start_ctid = X,
> > end_ctid=Y), and it will stay in L1 cache most of the time you process
> > the corresponding relation. and the relation itself will be smaller, and
> > index-only (actually index-only + lookup inside L1 cache) access can
> > happen, and so on .
> >
> > OTOH, if you load it in millions of small transactions, you can run
> > VACUUM FREEZE _on_ the visibility heap only, which will make all
> > visibility infoe look similar and thus RLE-compressable and again make
> > it fit in L1 cache, if you dont have lots of failed loads interleaved
> > with successful ones.
> >
> > > Bitmap index with index-only access does provide nearly all of the
> > > advantages of a column store from a speed standpoint BTW.  Even though
> > > Vertica is touting speed advantages - our parallel engine plus bitmap
> > index
> > > will crush them in benchmarks when they show up with real code.
> > >
> > > Meanwhile they're moving on to new ideas - I kid you not "Horizontica"
> > is
> > > Dr. Stonebraker's new idea :-)
> >
> > Sounds like a result of a marketroid brainstorming session :P
> >
> > > So - bottom line - some ideas from column store make sense, but it's not
> > a
> > > cure-all.
> > >
> > > > There is also a MonetDB/X100 project, which tries to make MonetOD
> > > > order(s) of magnitude faster by doing in-page compression in order to
> > > > get even more performance, see:
> > >
> > > Actually, the majority of the points made by the MonetDB team involve
> > > decreasing the abstractions in the processing path to improve the IPC
> > > (instructions per clock) efficiency of the executor.
> >
> > The X100 part was about doing in-page compression, so the efficiency of
> > disk to L1 cache pathway would increase. so for 1/2 compression the CPU
> > would get twice the data threoughput.
> >
> > > We are also planning to do this by operating on data in vectors of
> > projected
> > > rows in the executor, which will increase the IPC by reducing I-cache
> > misses
> > > and improving D-cache locality.  Tight loops will make a much bigger
> > > difference when long runs of data are the target operands.
> > >
> > > - Luke
> > >
> > >
> > >
> > > ---------------------------(end of broadcast)---------------------------
> > > TIP 7: You can help support the PostgreSQL project by donating at
> > >
> > >                 http://www.postgresql.org/about/donate
> >
> >

[ Attachment, skipping... ]

> 
> ---------------------------(end of broadcast)---------------------------
> TIP 2: Don't 'kill -9' the postmaster

-- 
  Bruce Momjian  <bruce(at)momjian(dot)us>        http://momjian.us
  EnterpriseDB                             http://enterprisedb.com

  + If your life is a hard drive, Christ can be your backup. +

In response to

pgsql-hackers by date

Next:From: Peter EisentrautDate: 2008-04-10 19:07:08
Subject: Re: Commit fest queue
Previous:From: Bruce MomjianDate: 2008-04-10 18:50:29
Subject: Re: Indexam interface proposal

pgsql-patches by date

Next:From: Andrew ChernowDate: 2008-04-10 19:39:35
Subject: Re: [PATCHES] libpq type system 0.9a
Previous:From: Andrew ChernowDate: 2008-04-10 18:32:04
Subject: Re: [PATCHES] libpq type system 0.9a

Privacy Policy | About PostgreSQL
Copyright © 1996-2014 The PostgreSQL Global Development Group