Gluster users - Jan 2016 - How to maintain HA using NFS clients if the NFS daemon process gets killed on a gluster node?

On 01/27/2016 09:39 PM, Kris Laib wrote:
> Hi all,
>
> We're getting ready to roll out Gluster using standard NFS from the
> clients, and CTDB and RRDNS to help facilitate HA.   I thought we were
> good to know, but recently had an issue where there wasn't enough
memory
> on one of the gluster nodes in a test cluster, and OOM killer took out
> the NFS daemon process.   Since there was still IP traffic between nodes
> and the gluster-based local CTDB mount for the lock file was intact,
> CTDB didn't kick in an initiate failover, and all clients connected to
For gluster-NFS, CTDB is typically configured to maintain high 
availability and I guess you have done the same. Could you check why 
CTDB hasn't initiated IP failover?

An alternative solution is to use nfs-ganesha [1][2] to provide NFS 
support for gluster volumes and can be configured to maintain HA using 
gluster CLI.

Thanks,
Soumya

[1] 
http://blog.gluster.org/2015/10/linux-scale-out-nfsv4-using-nfs-ganesha-and-glusterfs-one-step-at-a-time/

[2] 
http://gluster.readthedocs.org/en/latest/Administrator%20Guide/NFS-Ganesha%20GlusterFS%20Intergration/
(section# Using Highly Available Active-Active NFS-Ganesha And GlusterFS 
cli)
> the node where NFS was killed lost their connections.   We'll obviously
> fix the lack of memory, but going forward how can we protect against
> clients getting disconnected if the NFS daemon should be stopped for any
> reason?
>
> Our cluster is 3 nodes, 1 is a silent witness node to help with split
> brain, and the other 2 host the volumes with one brick per node, and 1x2
> replication.
>
> Is there something incorrect about my setup, or is this a known downfall
> to using standard NFS mounts with gluster?
>
> Thanks,
>
> Kris
>
>
>
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://www.gluster.org/mailman/listinfo/gluster-users
>

Soumya,

CTDB failover works great if the server crashes or the NIC is pulled, but I
don't believe there's anything in the CTDB setup that would cause it to
realize there is a problem if only the glusterfs process responsible for serving
NFS is killed but network connectivity with other CTDB nodes remains intact.    
If others are able to kill just the PID for the associated "NFS Server on
localhost" process and have CTDB issue a failover, I'd be very
interested to know how their setup differs from mine.

Thanks for the nfs-ganesha suggestion, I'm not very familiar with that
option and don't have enough time in my timeline to properly test it before
moving to production, but I will look into it further for a possible solution
down the road or if my deadline gets extended.   The FUSE client may be a good
option for us as well, but I can't seem to get speeds higher than 30 MB/s
using the Gluster FUSE client (I posted more details on that earlier today to
this group as well, looking for advice there).

-Kris

________________________________________
From: Soumya Koduri <skoduri at redhat.com>
Sent: Wednesday, January 27, 2016 8:15 PM
To: Kris Laib; gluster-users at gluster.org
Subject: Re: [Gluster-users] How to maintain HA using NFS clients if the NFS
daemon process gets killed on a gluster node?

On 01/27/2016 09:39 PM, Kris Laib wrote:
> Hi all,
>
> We're getting ready to roll out Gluster using standard NFS from the
> clients, and CTDB and RRDNS to help facilitate HA.   I thought we were
> good to know, but recently had an issue where there wasn't enough
memory
> on one of the gluster nodes in a test cluster, and OOM killer took out
> the NFS daemon process.   Since there was still IP traffic between nodes
> and the gluster-based local CTDB mount for the lock file was intact,
> CTDB didn't kick in an initiate failover, and all clients connected to
For gluster-NFS, CTDB is typically configured to maintain high
availability and I guess you have done the same. Could you check why
CTDB hasn't initiated IP failover?

An alternative solution is to use nfs-ganesha [1][2] to provide NFS
support for gluster volumes and can be configured to maintain HA using
gluster CLI.

Thanks,
Soumya

[1]
http://blog.gluster.org/2015/10/linux-scale-out-nfsv4-using-nfs-ganesha-and-glusterfs-one-step-at-a-time/

[2]
http://gluster.readthedocs.org/en/latest/Administrator%20Guide/NFS-Ganesha%20GlusterFS%20Intergration/
(section# Using Highly Available Active-Active NFS-Ganesha And GlusterFS
cli)
> the node where NFS was killed lost their connections.   We'll obviously
> fix the lack of memory, but going forward how can we protect against
> clients getting disconnected if the NFS daemon should be stopped for any
> reason?
>
> Our cluster is 3 nodes, 1 is a silent witness node to help with split
> brain, and the other 2 host the volumes with one brick per node, and 1x2
> replication.
>
> Is there something incorrect about my setup, or is this a known downfall
> to using standard NFS mounts with gluster?
>
> Thanks,
>
> Kris
>
>
>
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://www.gluster.org/mailman/listinfo/gluster-users
>