+++ This bug was initially created as a clone of Bug #1283956 +++

Description of problem:

I have a 3 node setup with 1 arbiter brick for every volume. Every volume contains a couple of big files with KVM disk images.

Every couple of minutes/seconds (probably depends on activity), a self-heal operation is triggered on one or more files on the volumes. During that time, there's no noticable loss of connectivity or anything like that.

How reproducible:

Run "gluster volume heal volume-name info" a couple of times and observe the output.


Actual results:

root@web-vm:~# gluster volume heal system_www1 info
Brick cluster-rep:/GFS/system/www1
/images/101/vm-101-disk-1.qcow2 - Possibly undergoing heal

Number of entries: 1

Brick web-rep:/GFS/system/www1
/images/101/vm-101-disk-1.qcow2 - Possibly undergoing heal

Number of entries: 1

Brick mail-rep:/GFS/system/www1
/images/101/vm-101-disk-1.qcow2 - Possibly undergoing heal

Number of entries: 1

Expected results:

Heal not being triggered without a reason.

Additional info:

Setting "cluster.self-heal-daemon" to "off" on the volumes does not change the behavior.

--- Additional comment from Adrian Gruntkowski on 2015-11-20 06:41 EST ---



--- Additional comment from Ravishankar N on 2015-11-20 07:07:51 EST ---

Had a quick look at one of the mount logs for the 'system_ww1 volume', i.e. glusterfs_cluster-vm/mnt-pve-system_www1.log.1 where I do see disconnects to the bricks.

#grep -rne "disconnected from" mnt-pve-system_www1.log.1|tail -n3
2177:[2015-11-19 15:58:32.687248] I [MSGID: 114018] [client.c:2042:client_rpc_notify] 0-system_www1-client-0: disconnected from system_www1-client-0. Client process will keep trying to connect to glusterd until brick's port is available
2283:[2015-11-19 15:58:43.486658] I [MSGID: 114018] [client.c:2042:client_rpc_notify] 0-system_www1-client-0: disconnected from system_www1-client-0. Client process will keep trying to connect to glusterd until brick's port is available
2385:[2015-11-19 15:58:43.557338] I [MSGID: 114018] [client.c:2042:client_rpc_notify] 0-system_www1-client-2: disconnected from system_www1-client-2. Client process will keep trying to connect to glusterd until brick's port is available

So it appears that there are network disconnects from the mount to the bricks. If I/O was happening during the disconnects, when the connection is re-established, self-heal woll get triggered.

Adrian, could you confirm like you said on IRC if it could be an issue with your firewall/ network? If yes, Ill close it as NOTABUG

--- Additional comment from Ravishankar N on 2015-11-20 07:11:14 EST ---

(In reply to Adrian Gruntkowski from comment #0)

> Setting "cluster.self-heal-daemon" to "off" on the volumes does not change
> the behavior.

Clients (mounts) can also trigger self-heals in addition to the self-heal daemon. If you want to disable client side heal, you need to set cluster.metadata-self-heal, cluster.data-self-heal and cluster.entry-self-heal to off.

--- Additional comment from Adrian Gruntkowski on 2015-11-20 08:06:01 EST ---

The entries that you mentioned have timestamps from yesterday. I did a couple of server restarts and was fiddling applying the patch and so forth. The logs look clean for today in that regard.

I have double checked the logs for interface flapping and firewall rules but everything seems fine. The pings on the interfaces dedicated to gluster between the nodes go through without any losses.

Ravishankar: Sure, I was changing that setting in the course of experiment that Pranith wanted me to do. Just mentioned it for completeness.

--- Additional comment from Pranith Kumar K on 2015-11-24 09:22:00 EST ---

hi Adrian,
      I looked at the pcap files and found nothing unusual. So I think we are left with trying to re-create the problem. Do you think we can come up with a way to recreate this problem consistently?

Pranith

--- Additional comment from Adrian Gruntkowski on 2015-11-24 09:30:06 EST ---

My setup is pretty basic, save for crossover configuration of 2 sets of volumes.
I have actually laid it out in the initial post on ML about the issue:

http://www.gluster.org/pipermail/gluster-users/2015-October/024078.html

For the time being, I'm rolling back to a 2-node setup. I will also try to setup a cluster with arbiter in a local test env on virtualbox based VMs.

Adrian

--- Additional comment from Pranith Kumar K on 2015-11-24 11:11:32 EST ---

Adrian,
     So you don't see this without Arbiter?

Pranith

--- Additional comment from Adrian Gruntkowski on 2015-11-25 04:00:23 EST ---


Yes, I see it only when in arbiter setup.

Adrian

--- Additional comment from Vijay Bellur on 2015-11-26 00:01:57 EST ---

REVIEW: http://review.gluster.org/12755 (cluster/afr: change data self-heal size check for arbiter) posted (#1) for review on master by Pranith Kumar Karampuri (pkarampu)

--- Additional comment from Pranith Kumar K on 2015-11-26 00:02:43 EST ---

Adrian,
    I was able to recreate this issue.

Steps to recreate:
1) Create a volume with arbiter, start the volume and mount the volume
2) On the mount point execute "dd if=/dev/zero of=a.txt"
3) While the command above is running, execute "gluster volume heal <volname> info" in a loop. We will see pending entries to be healed.

With the patch in https://bugzilla.redhat.com/show_bug.cgi?id=1283956#c9
I don't see the issue anymore. Let me know how your testing goes with this patch.

Pranith