[Lustre-discuss] MDS backup and restore fail

[Andreas Dilger]

On Apr 04, 2009  00:32 -0500, dan wrote:
> I was switching MDS servers, all new hardware.  I followed the backup and
> restore procedures in the manual and some tips from the list.  When it was
> all said and done the OSTs mounted up and I could mount a client just fine.
>  It seemed like everything was fine until I realized all the files on the
> system are zero length instead of 100's of GB.  What happened?

It sounds like you did not back up the extended attributes.  IIRC is a
step in the restore process that verifies that the attributes were
restored correctly.

> Background:
> 
> I used rsync -aXvv.  It was very very slow over ssh or native rsync so I
> mounted the old MDS partition via NFS on the new server - I think this
> might be the problem.  Do the extended attributes transfer over NFS?

I don't think there is support for this in NFS.

> Some files in /ROOT ended up being 80 MB instead of the normal 512 byte MDT
> files.  The original file system was 794 MB, the new one is 166 GB.
> 
> I can create new files that are normal.  Performance is good.

I would NOT continue creating new files until you have successfully
restored all of the attributes.  You can re-copy the xattrs by themselves
by mounting the old MDS with -t ldiskfs and using "getfattr" as documented,
and then "setfattr" on the new MDS.

> I'm currently running e2fsck to build the databases for lfsck.  Will that
> help or hurt this problem?  What's the best way to recover from this? 
> Thank you!

This will not help you, and would likely erase all of your data objects
if you run lfsck with that database.

Cheers, Andreas
--
Andreas Dilger
Sr. Staff Engineer, Lustre Group
Sun Microsystems of Canada, Inc.