Thread: conversion security update may have slowed our system?
Dear Gurus, I dunno if it should go to -perform or -bugs, so stay with the original message. As Tom wrote in http://archives.postgresql.org/pgsql-announce/2005-05/msg00001.php as well as at http://www.postgresql.org/about/news.315, I patched all our testing (7.4.6) and production databases (7.3.3) on Thursday. According to user reports, the production server slowed down noticeably on Friday and is almost unbearably slow today (Monday). My humble question is, could you think of any way this patch could affect overall speed, or is it just a fatal coincidence? System: Debian "Woody" @ DualXeon 2.4 1G 5x10krpm scsi hw raid 5 IBM server Kernel 2.4.30 (patched on Tuesday, 3 days before first syndroms TESTS: 1. As far as we could test without modifying the prod.db (i.e. issuing updates/inserts), this server makes both simple (select *) and joined queries in about twice the time of an Athlon 2k server on 7.4. VACUUM FULL VERBOSE ANALYZE and ANALYZE in itself did not help. 2. Set up another database, to see if it's in-db issue, but it made no difference. 3. Set up another database, based on a pre-patch dump (but I think it's irrelevant since I also patched template0, but it made no difference. Any ideas? Since I haven't checked it before, could you please provide an "undo" for a 7.3.3 database? (I assume it's a similar update, but don't know the original ACL values) Yours, -- G.
=?ISO-8859-2?Q?Sz=FBcs_G=E1bor?= <surrano@gmail.com> writes:
> According to user reports, the production server slowed down noticeably on
> Friday and is almost unbearably slow today (Monday). My humble question is,
> could you think of any way this patch could affect overall speed, or is it
> just a fatal coincidence?
I can't see that there could be any connection. You need to be looking
for other explanations, not an "undo".
regards, tom lane
On Mon, 2005-05-09 at 09:54 -0400, Tom Lane wrote: > =?ISO-8859-2?Q?Sz=FBcs_G=E1bor?= <surrano@gmail.com> writes: > > According to user reports, the production server slowed down noticeably on > > Friday and is almost unbearably slow today (Monday). My humble question is, > > could you think of any way this patch could affect overall speed, or is it > > just a fatal coincidence? > > I can't see that there could be any connection. You need to be looking > for other explanations, not an "undo". Szucs, Your original release was 7.3.3 and you have just moved to 7.3.9. Thats a large jump to have made, so many things are likely to have changed. It doesn't seem likely that it was the 7.3.8 -> 7.3.9 jump, but many others things have changed. You need to establish some means of measuring your system's performance other than user reports, so you can verify problems objectively. Tom, It seems reasonable to publish some backout information with every patch to say whether or not it is possible to downgrade should a problem be found. That could also be accompanied with a "has not been tested" caveat if need be. Maybe Szucs hasn't found one this time, but there is a possibility and many people don't have a long time to decide with a production system looking unstable. Best Regards, Simon Riggs
Simon Riggs <simon@2ndquadrant.com> writes:
> Your original release was 7.3.3 and you have just moved to 7.3.9.
Did he move to 7.3.9? I got the impression he'd only applied the
recommended catalog change to his existing installation. But if he
did, then you're right, it's more than likely got something to do
with some other change in the 7.3 series.
regards, tom lane
On Mon, 2005-05-09 at 14:10, Tom Lane wrote: > Simon Riggs <simon@2ndquadrant.com> writes: > > Your original release was 7.3.3 and you have just moved to 7.3.9. > > Did he move to 7.3.9? I got the impression he'd only applied the > recommended catalog change to his existing installation. But if he > did, then you're right, it's more than likely got something to do > with some other change in the 7.3 series. Most of the time I've seen this happen, the change has made the query planner do something differently, and explain analyze has shown the problem rather quickly.