Thread: PG over NFS tips

PG over NFS tips

From
Cott Lang
Date:
The higher-ups are attempting to force me to run Postgres over NFS at
least temporarily.

Despite giving me a queasy feeling and reading quite a bit of messages
advising against it, running Oracle over NFS with a NAS filer doesn't
seem to be unusual. Is there a reason PG would be more sensitive than
Oracle?

Anyone ever done this before in a production environment?

thanks!



Re: PG over NFS tips

From
Tom Lane
Date:
Cott Lang <cott@internetstaff.com> writes:
> The higher-ups are attempting to force me to run Postgres over NFS at
> least temporarily.

> Despite giving me a queasy feeling and reading quite a bit of messages
> advising against it, running Oracle over NFS with a NAS filer doesn't
> seem to be unusual. Is there a reason PG would be more sensitive than
> Oracle?

No --- the issues are not with Postgres per se but with the reliability
of your NFS setup.  On top of the not-infinite reliability of disk drive
hardware you now have to stack risk of failure of the NAS machine itself,
network problems, and misconfiguration problems (eg, ill-chosen mount
options).

The people who run Oracle over NFS successfully have usually paid top
dollar for quality NAS hardware and a network run by people who know
what they're doing.  Put PG into that same environment and it will work
just as well.  But put PG on a lesser-grade setup, run by not quite such
competent admins, and you're in for trouble.

            regards, tom lane

Re: PG over NFS tips

From
Cott Lang
Date:
On Thu, 2004-08-05 at 07:37, Tom Lane wrote:

> No --- the issues are not with Postgres per se but with the reliability
> of your NFS setup.  On top of the not-infinite reliability of disk drive
> hardware you now have to stack risk of failure of the NAS machine itself,
> network problems, and misconfiguration problems (eg, ill-chosen mount
> options).

All too well understood - although not by the higher-ups.

> The people who run Oracle over NFS successfully have usually paid top
> dollar for quality NAS hardware and a network run by people who know
> what they're doing.  Put PG into that same environment and it will work
> just as well.  But put PG on a lesser-grade setup, run by not quite such
> competent admins, and you're in for trouble.


This will be with an EMC Celerra - so it's at least top dollar. The jury
is still out on quality. :)

I was hoping to cable the machines directly to the NAS, but despite
promises to the contrary, on arrival it lacks enough ports for redundant
crossover connections.

I'm now planning on using the Linux bonding ethernet driver and
etherchannels for redundancy - redundant ethernet cards, links,
switches, and even NAS heads in the filer. I'm not sure how to setup an
NFS network any better, but I welcome any suggestions!

FWIW, I'm being forced to do this because EMC has not (so far) validated
using Opterons in 64bit mode attached to a Clariion SAN. Consider this a
big heads-up for anyone else out there thinking about using EMC +
Opteron.







Re: PG over NFS tips

From
Gaetano Mendola
Date:
Cott Lang wrote:

> The higher-ups are attempting to force me to run Postgres over NFS at
> least temporarily.
>
> Despite giving me a queasy feeling and reading quite a bit of messages
> advising against it, running Oracle over NFS with a NAS filer doesn't
> seem to be unusual. Is there a reason PG would be more sensitive than
> Oracle?
>
> Anyone ever done this before in a production environment?
>
> thanks!

Do you trust your data to a udp connection ?
We had problem in copying big files ( 1.9GB ) in a mounted NFS partition
and now we prefer to not use it anymore for our data.



Regards
Gaetano Mendola





Re: PG over NFS tips

From
Marco Colombo
Date:
Gaetano Mendola wrote:
> Do you trust your data to a udp connection ?

- Last time I checked, UDP was connectionless. :)
- NFS runs over TCP, too.
- TCP isn't any better than UDP when it comes to data safety. Either the
   app does its own checksumming, or you trust the datalink layer.

> We had problem in copying big files ( 1.9GB ) in a mounted NFS partition
> and now we prefer to not use it anymore for our data.

I think some (old) versions of NFS and/or (old) Linux kernels have issues
with files >2GB. That 1.9GB is suspiciously close to a 2^31 file offset
limit. I have a 2.2.xx based NFS server here, and of course I can't copy
any file bigger than 2GB on it (since the filesystem on the server doesn't
support that).

I think PostgreSQL handles that well, but I dunno how it detects whether
the filesystem supports 64 bit offsets or not.

.TM.
--
       ____/  ____/   /
      /      /       /            Marco Colombo
     ___/  ___  /   /              Technical Manager
    /          /   /             ESI s.r.l.
  _____/ _____/  _/               Colombo@ESI.it