[Gluster-devel] split brain: how should it be cured?

[Emmanuel Dreyfus]

On Wed, Jun 20, 2012 at 05:35:36AM -0400, Pranith Kumar Karampuri wrote:
> Procedure to reset split-brain is to remove one of the files. 
> 3.3 release onwards we need to also delete the file from gfid-backend.
> if the file is a.txt with gfid 5c71a049-a424-480e-9e69-2cdfd7a0a7a0
> remove the file a.txt
> remove the file <brick-path>/.glusterfs/5c/71/5c71a049-a424-480e-9e69-2cdfd7a0a7a0
> the gfid path is a hardlink to a.txt

glusterfs 3.3 behavior is really odd. It was possible to read the
offending file, and after a while, it started  returning EIO again.
Removing the file and its gfid hardlink seems to workaround the
problem, though.

But I still do not understand how they are considered in metadata split brain: 
all metadata were the same:

On brick1:
  File: "/export/wd3a/manu/netbsd/usr/src/tools/mktemp/Makefile"
  Size: 165          FileType: Regular File
  Mode: (0644/-rw-r--r--)         Uid: (  500/    manu)  Gid: (  500/    manu)
Device: 0,16   Inode: 55748693    Links: 2
Access: Wed Jun 20 13:24:31 2012
Modify: Sun Dec  8 21:20:03 2002
Change: Wed Jun 20 11:31:28 2012

trusted.afr.pfs-client-0 00 00 00 00 00 00 00 00 00 00 00 00
trusted.afr.pfs-client-1 00 00 00 00 00 00 00 01 00 00 00 00

On brick2:
  File: "/export/wd3a/manu/netbsd/usr/src/tools/mktemp/Makefile"
  Size: 165          FileType: Regular File
  Mode: (0644/-rw-r--r--)         Uid: (  500/    manu)  Gid: (  500/    manu)
Device: 0,24   Inode: 25441937    Links: 2
Access: Wed Jun 20 13:24:31 2012
Modify: Sun Dec  8 21:20:03 2002
Change: Wed Jun 20 11:31:28 2012

trusted.afr.pfs-client-0 00 00 00 00 00 00 00 03 00 00 00 00
trusted.afr.pfs-client-1 00 00 00 00 00 00 00 00 00 00 00 00
-- 
Emmanuel Dreyfus
manu at netbsd.org