On Thursday 26 February 2004 12:09 pm, Jeffrey W. Baker wrote:
> On Thu, 2004-02-26 at 11:46, Holger Marzen wrote:
> > On Thu, 26 Feb 2004, Jeffrey W. Baker wrote:
> > > We are using PostgreSQL with the database and xlogs on
> > > (separate) XFS volumes under Linux 2.4.25. We are simply
> > > curious to hear your experiences with this combination, if you
> > > are using it. In only two days of heavy activity, we've
> > > already been able to corrupt one database. We've also seen XFS
> > > panic because of inconsistent in-memory metadata. Frankly we
> > > don't have the highest confidence.
> >
> > I am afraid that xfs in that kernel or your hardware is buggy
> > (probably RAM). A 24h run of memtest86 wouldn't be bad.
> >
> > Since PostgreSQL uses the operating system's calls for file
> > operations as any other program does, it's most probably no
> > PostgreSQL issue.
>
> I don't see why not. PostgreSQL could easily have a bug that swaps
> a buffer somewhere, resulting in a corrupt table. That we see this
> only on the INSERT path and not the COMMIT path also seems to point
> towards Pg.
>
> Anyway, you didn't mention XFS. Do you have experience using it
> beneath Postgres?
I do. It's great.
One of my PostgreSQL machines uses XFS with everything (data, xlogs,
etc.) on the same XFS partition. Machine uptime: 247 days. PostgreSQL
uptime 74 days (restarted while testing reconfiguration options).
PostgreSQL crashes: 0.
As the good doc says, "when you hear hoofbeats think horses, not
zebras." Don't chase some hypothetical theoretical possible problem
when it's known that bad RAM causes problems, especially when your
error message seems to point toward RAM trouble.
I give all my new machines a good test with memtest86 or the newer
memtest86+ before installation. Just two weeks ago I got a machine
that looked good after the first couple of passes but when I went to
do a final check and install it a week and 1000 passes later it
reported 6 errors in one memory location - a frustrating random crash
just waiting to happen.
Cheers,
Steve
