SUMMARY: ADVfs data journaling and SYNC writes

From: Lucien HERCAUD <Lucien_HERCAUD_at_paribas.com>
Date: Wed, 27 Nov 1996 11:28:44 +0000

     Hi,
     
     and thanks for replys to (so far):
     
>>> webster_at_europa.mdc.com:
     ---------------------------
>>> I agree that the chfile manual page is misleading, but I have never
>>> seen a sane filesystem implementation that can guarantee application
>>> file stability...
     
>>> ohad_at_comport.com :
     ----------------------
>>> Sybase does not support any unix file system! ...
>>> ... Oracle supports running on ADVFS with integrity!
     
     My own conclusions:
     
     As far as I know, Sybase used to support running on top of JFS onto AIX.
     Does this mean Sybase did a special AIX implementation ?
     
     I used to test Sybase on ADVfs with DATA LOGGING == ON and deliberately
     provoqued tens of crashes and power outages. No problem so far ! BUT when
     data logging was off, the database check reported major corruptions. In
     addition, Sybase stated that if DEC supports data logging onto ADVfs, they
     "support" the FS implementation. Sybase assumed data logging to guarantee
     the fact that a successfull write returned to application meanned that
     data was on stable media.
     
     So, we used this ADVfs implementation with a life system. But then my
     system started to crash and DEC UNIX engineering stated there is a problem
     (which they did not know about until me escalating it) with the DATA
     LOGGING being turned on for ordinary files. I know they will fix this
     problem in the future; but, in the meanwhile, they tried to prove to me
     that "chfile -l on" brings no benefit - other than statistic ! (as you
     could read from my mail).
     
     So, yes, I'll move all Sybase installations to raw device (or raw LSM),
     but this may also mean we no longer use DIGITAL UNIX for Sybase.
     
     Cheers,
                                /lh

Full responses below:
=====================

hello there,
     
   Sybase does not support any unix file system! the only way to get
   reliebility from Sybase s to run all your disks raw. I just finished doing
   3 large Sybase installes. If you have Sybase running on a file system
   and the system crashes - you will have to restore from tape! The three
   things that we hope for Sybase in the future are the FS support, large
   block size, and better read-ahead.
     
     
   Oracle supports running on ADVFS with integrity! I have been running Oracle
   on D.U. + ADVFS for about 2.5 years, and never had a problem! Even when
   one of the machines crashed - Oracle came back without any problems.
     
     
   Just like archive loggin enables Oracle to keep all the transactions
   since last backup - So does the ADVFS logging. It saves all changes to
   the file - so the having the old data on disk and having all the loggin
   information you then can recreate the actual new data. This is also how
   you are able to create a clone files set and get a good backup of it!
     
     
   Good luck,
     
     Ohad
     
     ----------------------------------------------------------
     
     Lucien,
     
     I agree that the chfile manual page is misleading, but I have never seen a
     sane filesystem implementation that can gaurantee application file
     stability.
     
     The promise of journaling filesystems has always been that the system
     would ALWAYS come up with a stable filesystem. Writes that had not been
     flushed to the logs or had not finished being written to the logs would be
     lost.
     
     The chfile logging option seems to just be a way of instructing ADVFS to
     write more, smaller logs to the journal, more often. While it would be
     possible
     to require that the log had been written to the journal before the write()
     returned, I think you would find that performance would go right down the
     drain.
     
     Besides that, is Sybase willing to confirm that all of the needed updates
     to retain application stability are done as part of a single write() call?
      I don't think that they will. Even if all writes are un-buffered, I'd
     have to belive that there are points where a system failure would render
     databases or indexes unstable. (Ok, maybe in a PURE TP envirionment, you
     would get the logical equivelant of a journaling filesystem for the
     DB....) Is this the claim
     Sybase is making with raw partition use?
     
     In any event, if you:
     
     1. Physically secure the server.
     2. Connect the server to a good UPS.
     3. Use redundant power supplies in the server.
     4. Use RAID-0 or RAID-5 (with no controler caching or battery backed
     caching). 5. Use ECC RAM.
     
     I can't think of much that would manage to munge the system (other than
     something
     that the users or sysadmin does via software (intended or not)) that these
     steps wouldn't cover. If you make the assumtion that the software is
     running, everything
     else is a matter of providing a safe environment for the system, where you
     can gaurantee that it always has time to flush its buffers.
     
     BTW - CC:Mail is encoding your message as an enclosure, rather than as
     text encoded
     mesage text. You may want to look into this, or not.
     
     Just my $0.02's worth,
     
     Tom
Received on Wed Nov 27 1996 - 12:02:11 NZDT

This archive was generated by hypermail 2.4.0 : Wed Nov 08 2023 - 11:53:47 NZDT