[Gluster-devel] solutions for split brain situation

[Anand Avati]

>  When reading the GlusterFS replication description - it seemed like they
> discussed what happens if a node goes down *during* the unlink(), but they
> didn't specifically cover what happens if the node is down before and after
> the unlink(), but then brought back up some time later. One interpretation
> suggested to me that the journal would not hang around, and the file would
> be treated as gone from all "up" nodes, and then when the later node is
> brought back up, the file shows up as existing on only one node, and the
> self-heal would kick in and replicate the file as if it still exists. I
> didn't find the time to test this exact scenario yet, though...

>From a top level - if the unlink() system call has returned success
back to the app, then your file is assured to be deleted no matter at
what stage any server went down.

If a server is already down even before of the unlink() call was
initiated by the app, then the file is deleted when the server comes
back up.

If a server is brough down in the middle of an unlink transaction, no
matter at what stage of the transaction (out of the 5 stages - lock,
journal, act, unjournal, unlock) then the file is assured to get
deleted (if it the server went down before the 'act's tage) when the
server comes back - provided the transaction has proceeded to
completion on atleast one server.

Avati