Thread: wholly / partially table read into shared buffer

Hello!
I am newbie in Postgres, but previously worked with Oracle. Now I am puzzled with heap_blks_hit/heap_blks_read ratio. I
makeselect from table (whose size is larger than shared_buffer) and saw that it was wholly read from disk after second,
thirdand so on runs. Isn't is wrong that RDBMS tries to cache table data blocks at least PARTIALLY in the case of
insufficientcache buffer? 
WBW&R, Grigory Titov.

On Thu, 2007-09-20 at 19:07 +0400, Григорий Титов wrote:

> I am newbie in Postgres, but previously worked with Oracle. Now I am
> puzzled with heap_blks_hit/heap_blks_read ratio. I make select from
> table (whose size is larger than shared_buffer) and saw that it was
> wholly read from disk after second, third and so on runs. Isn't is
> wrong that RDBMS tries to cache table data blocks at least PARTIALLY
> in the case of insufficient cache buffer?

It does cache partially, but its always the wrong part of the table.

BTQ, Oracle does that too...

--
  Simon Riggs
  2ndQuadrant  http://www.2ndQuadrant.com

On 24 Wrz, 23:59, si...@2ndquadrant.com (Simon Riggs) wrote:
> On Thu, 2007-09-20 at 19:07 +0400,                wrote:
> > I am newbie in Postgres, but previously worked with Oracle. Now I am
> > puzzled with heap_blks_hit/heap_blks_read ratio. I make select from
> > table (whose size is larger than shared_buffer) and saw that it was
> > wholly read from disk after second, third and so on runs. Isn't is
> > wrong that RDBMS tries to cache table data blocks at least PARTIALLY
> > in the case of insufficient cache buffer?
>
> It does cache partially, but its always the wrong part of the table.
>
Let's say - you have very big table. Even if you are querying data
with index -
table read will be done... If your queries are quite randomly hitting
table pages -
you *will* see that effect. You can think about partitions - idea
looks good -
you can save some IO bandwidth paying with processor time. Partitions
can
help if you do seq-scans, but... with index scans and heap_blks_hit -
does not.
Instead you can prepare 'index cluster' a specially when there are not
too many
writes...

--
Regards,
  Micha  Zaborowski (TeXXaS)