[Gluster-users] rogue file

Jeff Darcy jdarcy at redhat.com
Tue Jul 9 18:26:00 UTC 2013

On 07/09/2013 02:01 PM, Eco Willson wrote:
> Thinking about these scenarios after the file has been written:
> A. No one ever looks at the file
> Nothing will happen in this case
> B. A self-heal runs
> If the file has not had xattrs appended to it, self-heal will skip the file
> C. A directory read gets done on that sub-directory
> ls by itself will not append the xattrs, but a `stat` of the file will
> D. Someone actually tries to read the file (like if it was called readme.txt
> or some other common name)
> A stat or vi in this case will append the xattrs.
> In 3.4beta4 it appears even an ls on the file will work.

That seems accurate in the case where the "rogue" file is the only one with 
that path.  However, if another brick within the same replica set has a file 
with that path, something different will happen.  Specifically, the "rogue" 
file will be observed to have no GFID, which will cause operations to fail 
because of a GFID mismatch.  Technically that's not split brain, but the 
practical implications are the same.

More information about the Gluster-users mailing list