[Gluster-devel] AFR arbiter volumes

[Ravishankar N]

Sending out this mail for awareness/ feedback.
-----------------------------------------------------------------------------
*What:**
*Since glusterfs-3.7,  AFR supports creation of arbiter volumes. These 
are a special type of replica 3 gluster volume where the 3rd brick  is 
(always) configured as an arbiter node.What this means is that the 3rd 
brick will store only the file name and metadata (including gluster 
xattrs), but does not contain any data. Arbiter volumes prevent 
split-brains and consumes lesser space than a normal replica 3 volume 
and provides better consistency and availability than a replica 2 volume.

*How:**
*You can create an arbiter volume with the following command:
/
gluster volume create <VOLNAME> replica 3 arbiter 1 host1:brick1 
host2:brick2 host3:brick3/

Note that the syntax is similar to creating a normal replica 3 volume 
with the exception of the /arbiter 1/ keyword. As seen in the command 
above, the only permissible values for the replica count and arbiter 
count are 3 and 1 respectively. Also, the 3rd brick is always chosen as 
the arbiter brick and it is currently not configurable to have any other 
brick as the arbiter.

*Client/ Mount behaviour:**
*By default, client quorum (cluster.quorum-type) is set to auto for a 
replica 3 volume (including arbiter volumes) when it is created; i.e. at 
least 2 bricks need to be up to satisfy quorum and to allow writes. This 
setting is not to be changed for arbiter volumes also. Additionally, the 
arbiter volume has some additional checks to prevent files from ending 
up in split-brain:

     * Clients take full file locks when writing to a file as opposed to 
range locks in a normal replica 3 volume.

     * If 2 bricks are up and if one of them is the arbiter (i.e. the 
3rd brick) and it blames the other up brick, then all FOPS will fail 
with ENOTCONN (Transport endpoint is not connected). If the arbiter 
doesn't blame the other brick, FOPS will be allowed to proceed. 
'Blaming' here is w.r.t the values of AFR changelog extended attributes.

     * If 2 bricks are up and the arbiter is down, then FOPS will be 
allowed.

     * In all cases, if there is only one source before the FOP is 
initiated and if the FOP fails on that source, the application will 
receive ENOTCONN.

Note: It is possible to see if a replica 3 volume has arbiter 
configuration from the mount point. 
If/$mount_point/.meta/graphs/active/$V0-replicate-0/options/arbiter-count/ 
exists and its value is 1, then it is an arbiter volume. Also the client 
volume graph will have arbiter-count as a xlator option for AFR translators.

*Self-heal daemon behaviour:*

Since the arbiter brick does not store any data for the files, it cannot 
be used as a source for data self-heal. For example if there are 2 
source bricks B2 and B3 (B3 being arbiter brick) and B2 is down, then 
data-self-heal will not happen from B3 to sink brick B1, and will be 
pending until B2 comes up and heal can happen from it. Note that 
metadata and entry self-heals can still happen from B3 if it is one of 
the sources.

-----------------------------------------------------------------------------

Please provide feedback if you have tried it out.
*If you ever encounter a split-brain while using the arbiter volume, it 
is a BUG - do report!*
We have had users asking for a way to convert existing replica 2 volumes 
to arbiter volumes- this is definitely in our to-do list, in addition to 
some performance optimizations.

Thanks,
Ravi
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.gluster.org/pipermail/gluster-devel/attachments/20150909/b4da4b7e/attachment.html>