While the nfs protocol is stateless and thus should handle failing-over,
the procedures themselves are synchronous. Meaning, I am not sure
how a nfs client will handle getting a ok for some metadata change
(mkdir, etc) just before a server dies and is recovered by another node.
If the op did not make it to the journal, it would be a null op. But the
nfs client would not know that as the server has failed-over.
This is a qs for nfs.

What is the stack of the nfs clients? As in, what are they waiting on?

Luis Freitas wrote:
> Hi list,
>
>    I need to implement a High available NFS server. Since we already have 
> OCFS2 here for RAC, and already have a virtual IP on the RAC server that 
> failovers automatically to the other node, it seems a natural choice to use 
> it too for our NFS needs. We are using OCFS2 1.2. (Upgrade to 1.4 is not on 
> our current plans)
>
>    We did a preliminary failover test, and the client that mounts the 
> filesystem (Actually a solaris box) doesnt like the failover. We expect some 
> errors and minor data loss and can tolerate them as a transient condition, 
> but the problem is that the mounted filesystem on the client becomes useless 
> until we umount and remount it again.
>
>    I suspect that NFS uses inode numbers on underlying filesystem to create 
> "handles" that it passes on to clients, but I am not sure on how this is done.
>
>    Anyone know if we can achieve a failover without needing to remount he nfs 
> share on the clients? Any special options are needed mounting the OCFS2 
> filesystem and also for exporting it as NFS, or on the client?
>
> Best Regards,
> Luis
>   

_______________________________________________
Ocfs2-users mailing list
[email protected]
http://oss.oracle.com/mailman/listinfo/ocfs2-users

Reply via email to