[Gluster-users] Hundreds of duplicate files

Sun Dec 21 02:02:35 UTC 2014

Hi everyone,

We have a distributed Gluster volume on five bricks over two servers (first server running gluster 3.4.2, second server running gluster 3.5.1, both running Fedora 20)
Starting last week, doing a file listing on the mounted volume shows many files with the same name appearing twice (and they are listed with the same inode). Doing a search for these files, I have found 290,000 of them!!

If I do a listing of these files on the bricks themselves, it looks like most are link files (du will show the file on the first server as 0 bytes, and the sticky bit set). The file is fine on the second server. Unfortunately, running "getfattr -m - -e hex -d" on the file shows NO gluster-related attributes and I believe this is why both files appear in the listing. The files cannot be read by any programs as it is trying to read the link file. I assume the metadata became corrupted. This is a production server so we really need to know:

1. How did this happen, and how can we prevent it going forward? There was a server crash a week ago and I believe that was the cause.
2. How can we heal the Gluster volume/bricks and link files. If there is some straightforward way of restoring the link file pointer I can write a script to do it, obviously doing this manually will be impossible.

Thanks very much for any and all help - much appreciated!

Regards,
Tom

On Wed, Dec 17, 2014 at 4:07 AM, <tbenzvi at 3vgeomatics.com> wrote:
> Hi everyone, we have noticed some extremely odd behaviour with our
 > distributed Gluster volume where duplicate files (same name, same or
 > different content) are being created and stored on multiple bricks. The only
 > consistent clue is that one of the duplicate files has the sticky bit set. I
 > am hoping someone will be able to shed some light on why this is happening
 > and how we can restore the volume as there appear to be hundreds of such
 > files. I will try to provide as much pertinent information as I can.
 >
 > We have a 130TB Gluster volume consisting of two 20TB bricks on server1, and
 > three 40TB bricks on a server2 which were added at a later date (and
 > rebalancing was done). The volume is mounted on server1, and accessed only
 > through this server but by many users. Both servers went down due to power
 > loss several days ago after which this problem was first noticed. We ran a
 > rebalance command on the volumes, this has not fixed the problem.
 >
 >
 > Gluster volume info:
 > Volume Name: safari
 > Type: Distribute
 > Volume ID: d48d0e6b-4389-4c2c-8fd1-cd2854121eda
 > Status: Started
 > Number of Bricks: 5
 > Transport-type: tcp
 > Bricks:
 > Brick1: server1:/data/glusterfs/safari/brick00/brick
 > Brick2: server1:/data/glusterfs/safari/brick01/brick
 > Brick3: server2:/data/glusterfs/safari/brick02/brick
 > Brick4: server2:/data/glusterfs/safari/brick03/brick
 > Brick5: server2:/data/glusterfs/safari/brick04/brick
 >
 >
 > Size information:
 > /dev/sdc 37T 16T 22T 42% /data/glusterfs/safari/brick02
 > /dev/sdd 37T 16T 22T 42% /data/glusterfs/safari/brick03
 > /dev/sde 37T 17T 21T 45% /data/glusterfs/safari/brick04
 > /dev/md126 11T 7.7T 2.8T 74% /data/glusterfs/safari/brick00
 > /dev/md124 11T 8.0T 2.5T 77% /data/glusterfs/safari/brick01
 > server2:/safari 130T 63T 68T 48% /sar
 >
 >
 > Example 1:
 > -Two files with the same name exist in one directory
 > -They have different contents and attributes
 > -A file listing on the mounted volume shows the same inode
 > -The newer file has sticky bit set
 > -Neither file is corrupted, they can both be viewed by using the absolute
 > path (on the bricks)
 >
 > File listing on the mounted volume
 > 13036730497538635177 -rw-rw-r-T 1 jon users 924 Dec 15 10:42 RSLC_tab
 > 13036730497538635177 -rw-rw-r-- 1 jon users 418 Mar 18 2013 RSLC_tab
 >
 > Listing of the files on the bricks:
 > 8925798411 -rw-rw-r-T+ 2 jon users 924 Dec 15 10:42
 > /data/glusterfs/safari/brick00/brick/complete/shm/rs2/ottawa/mf6_asc/stack_org/RSLC_tab
 > 51541886672 -rw-rw-r--+ 2 1002 users 418 Mar 18 2013
 > /data/glusterfs/safari/brick02/brick/complete/shm/rs2/ottawa/mf6_asc/stack_org/RSLC_tab
 >
 >
 > Example 2:
 > -Two files with the same name exist in one directory
 > -They have the same content and attributes
 > -No sticky bit is set when looking at file listing on the mounted volume
 > -Sticky bit is set for one while when looking at file listing on the bricks
 > -Files are corrupted
 >
 > File listing on the mounted volume:
 > 13012555852904096080 -rw-rw-r-- 1 tom users 2393848 Dec 8 2013
 > ifg_lr/20130226_20130813.diff.phi.ras
 > 13012555852904096080 -rw-rw-r-- 1 tom users 2393848 Dec 8 2013
 > ifg_lr/20130226_20130813.diff.phi.ras
 >
 > Listing of the files on the bricks:
 > 17058578 -rw-rw-r-T+ 2 tom users 2393848 Dec 13 17:11
 > /data/glusterfs/safari/brick00/brick/rsc/rs2/calgary/u22_dsc/stack_org/ifg_lr/20130226_20130813.diff.phi.ras
 > 57986922129 -rw-rw-r--+ 2 1010 users 2393848 Dec 8 2013
 > /data/glusterfs/safari/brick02/brick/rsc/rs2/calgary/u22_dsc/stack_org/ifg_lr/20130226_20130813.diff.phi.ras
 >
 >
 > Additionally, only some files in this directory are duplicated. The
 > duplicated files are corrupted (can not be viewed as Raster images: the
 > original file type)
 > The files which are not duplicated are not corrupted.
 >
 > File command: (notice duplicate and singleton files)
 > ifg_lr/20091021_20100218.diff.phi.ras: Sun raster image data, 1208 x 1981,
 > 8-bit, RGB colormap
 > ifg_lr/20091021_20101016.diff.phi.ras: data
 > ifg_lr/20091021_20101016.diff.phi.ras: data
 > ifg_lr/20091021_20101109.diff.phi.ras: Sun raster image data, 1208 x 1981,
 > 8-bit, RGB colormap
 > ifg_lr/20091021_20101203.diff.phi.ras: Sun raster image data, 1208 x 1981,
 > 8-bit, RGB colormap
 > ifg_lr/20091021_20101227.diff.phi.ras: Sun raster image data, 1208 x 1981,
 > 8-bit, RGB colormap
 > ifg_lr/20091021_20110120.diff.phi.ras: Sun raster image data, 1208 x 1981,
 > 8-bit, RGB colormap
 > ifg_lr/20091021_20110213.diff.phi.ras: data
 > ifg_lr/20091021_20110213.diff.phi.ras: data
 > ifg_lr/20091021_20110309.diff.phi.ras: data
 > ifg_lr/20091021_20110309.diff.phi.ras: sticky data
 > ifg_lr/20091021_20110402.diff.phi.ras: Sun raster image data, 1208 x 1981,
 > 8-bit, RGB colormap
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.gluster.org/pipermail/gluster-users/attachments/20141220/b7051f6f/attachment.html>