[Linux-cluster] fencing node and filesystem

[brem belguebli]

Hi,
What you have to look at before making assumptions on whether or not
there will be corruption is how the application (in this case Oracle)
opens its files (datafiles, redo logs etc...).

>From what I remmember, Oracle defaults to O_DSYNC which tells to VFS
to bypass the cache and write directly to the physical disks.

This can be verifiable with lsof (+fg flag if it exists).

Brem

2010/2/9 Rajagopal Swaminathan <raju.rajsand at gmail.com>:
> Greetings,
>
> On Mon, Feb 8, 2010 at 8:19 PM, Bob Peterson <rpeterso at redhat.com> wrote:
>> ----- "Muhammad Ammad Shah" <mammadshah at hotmail.com> wrote:
>> | when a node(x) fenced by another node(y), will it corrupts the file
>> | system or no? and if the database was running and node(x) fenced by
>> | fenced node(y) will it corrupts the database or whats about the
>> | transaction ?
>>
>> With GFS and GFS2 there should be no metadata corruption, in theory.
>> When quorum is regained (or if it's not lost) the fenced
>> node's journal is replayed so the file system metadata should
>> remain intact.  That does not mean that data won't be lost.
>> If the node was fenced while it had data in its cache, that data
>> may be lost.  In critical situations, you can mark files and/or
>> directories with a flag that makes the data journaled as well,
>> so the data should be protected.  Of course, that has a performance
>> penalty.  There's a small section about this in the GFS faq.
>
> Whatever Rob says is absolutely correct.
>
> Having said that, IIRC DBs like Oracle writes to disks every 3 seconds
> and Linux flushes to disk every five seconds anyways.
>
> So, one can say that the the data loss window could be at most 5
> seconds. Am I right?
>
> Thanks and regards,
>
> Rajagopal
>
> --
> Linux-cluster mailing list
> Linux-cluster at redhat.com
> https://www.redhat.com/mailman/listinfo/linux-cluster
>