Re: Can postgresql store its data on raw device now? - Mailing list pgsql-general

From Chris Browne
Subject Re: Can postgresql store its data on raw device now?
Date
Msg-id 871vptfkpl.fsf@dba2.int.libertyrms.com
Whole thread Raw
In response to Can postgresql store its data on raw device now?  (Lizzy M <lizzymy@gmail.com>)
List pgsql-general
lizzymy@gmail.com (Lizzy M) writes:
>     I have an problem: can postgresql store its data on the raw disks now?
>
>     I have checked the mail list and manual, but haven�t found the
> answer. In some early mails, they mentioned pg didn�t support this
> character. But how about now? Raw disk may reduce the risks brought by
> file system, so I want to use it to store my data.

No, this has not changed, and if anything, it is *less* likely to
change than ever.

Twenty years ago, filesystems were sufficiently fragile, unreliable
and slow that DBMS vendors found it worthwhile to avoid the risks by,
in effect, implementing their own filesystems as part of the DBMS.

- Since then, OS vendors have found it to be in their interests to
  improve both the robustness and performance of their filesystems,
  which has substantially diminished the would-be benefits.

- Compound this with the factor that implementing your own filesystem
  adds substantially to the complexity of managing and deploying the
  DBMS.

The two really interesting filesystems these days are ZFS and OCFS:
  http://en.wikipedia.org/wiki/ZFS
  http://en.wikipedia.org/wiki/OCFS

It seems much more useful to allow the OS vendors to improve their
filesystems, which helps us, than to use raw partitions, which
requires:

 a) Creating a portable framework for OS-agnostic access to raw
    partitions

    (because Linux != FreeBSD != Solaris != Windows NT != AIX != ...)

 b) Implementing our own filesystem atop raw partitions

 c) Implementing our own filesystem management tools (e.g. - like cp,
    rm, ..., since our OS tools won't work)

As an aggregate, this would add an extraordinary amount of work, and
with the improvements that have taken place lately, it is not evident
that this would buy us *any* of:
 a) Improved performance
 b) Improved reliability

Indeed, I think we could reasonably expect for there to be a
significant period of time during which performance and reliability
would be *degraded* by such an effort.
--
let name="cbbrowne" and tld="linuxdatabases.info" in name ^ "@" ^ tld;;
http://cbbrowne.com/info/rdbms.html
"I'm guilty of a lot of things, but I didn't ever do that.''
-- Bill Clinton, on why he had astroturf lining the back of his pickup
   truck
[In fact, he _DID_ do this, thus making life creepier than fiction...]

pgsql-general by date:

Previous
From: "Octavio Alvarez"
Date:
Subject: Re: limit table to one row
Next
From: "Chris Spotts"
Date:
Subject: Re: type cast in index