Re: Database on the file system

Guys,guys, please! fsck looses files only when they have grown and the growth hasn't reached the disk's master tables before the crash! Otherwise, fsck does not "loose" files. And what does ORACLE use for storing its data? Fixed length files in most cases anyway. Certainly in the case of the redo logs. So if system crashes, no redo log file lost anywhere! Simple.

Also, note that most modern file systems accept some form of "write-through:, ie, the fs lets the software write the file and returns only when the write is finished, instead of the old "I'll write it later but consider it done" technology.
ORACLE uses this kind of stuff for the redo log files. Veritas works slightly different than this, but the end result is the same.

Raw just bypasses the fs cache and all the other fs rigmarole that goes with it and lets ORACLE "write to the bone", so to speak. In theory, can be faster. But you need to have the ORACLE cache ready to replace the speed lost by not using the fs cache. And a few other less better known ORACLE caches as well.

In simple terms, the above describes MOL what happens. There is actually a lot of meat to go through in a subject like this, but this pretty much sums it up.

--
Cheers
Nuno Souto
nsouto_at_nsw.bigpond.net.au.nospam
Is there a nospam domain?
http://www.users.bigpond.net.au/the_Den Bass Chorng <bass_at_octel.com> wrote in message news:7ku4ne$kd8$1_at_news.eng.octel.com...
> : If I choose to put redolog files and database files on the file system,
> : instead of raw devices and when I meet the system crash trouble,
> : I wonder why Oracle can completely make good to recover all database
> : files by redolog files.
>
> : I mean, as you may know, data-writing to all files on the file system
> : is generally kept once on kernel cash area so-called the file system
> : buffer and contents in the buffer are periodically flushed to the
> : physical disks by a syncer daemon.
> : Considering that system crash happen perior to buffer flush, I guess
> : some redolog files may be lost in next fsck phase. (Especially in
> : high-transaction environment, redolog file switching is also high.)
> : In case of loosing redolog files application no longer recover its
> : database, I think. It means it may lost some (many?) transactions.
>
> : I wonder why really safe way to put oracle files on the file systems,
> : instead of raw devices...
>
> Interesting point of view. I dont have an answer, but like
> to add another question. Isn't this problem still true
> even when you use raw ? Because when OS 'thinks' a write
> is commited, it actually has to go thru controller which
> has buffer and disk which also has buffer, so you can
> still lose transaction if a power outage happens at
> the right time. Bear in mind, it's a long way before
> it is finally physically writen to disk, unless you use
> subsystem like EMC which has battery powered cache. Loss
> of commited data is always possible, although chance is slim.
>
> Among all the debates between raw and cooked file system
> I have never heard any discussion on this respect. But
> all I know is we use filesystem for over 45 instances
> for years and have gone thru crashes and sudden reboots
> ( while transaction is happening ), the loss of transaction
> in redo has never happened.
Received on Fri Jun 25 1999 - 07:48:57 CDT