Hi,

We don't run mesos 1.11 but we use Zookeeper with dynamic reconfiguration 
capability without any issue for Mesos 1.9. The only thing that should be 
handled carefully is the addition/removal of Zookeeper members when using 
dynamic reconf feature.

What do you mean by "mesos-master can handle a dynamic reconfiguration of the 
zk ensemble" ? To my understanding, Mesos will only connect to ZK to elect a 
leader through ZK primitives; I don't think there is a correlation with how ZK 
members are set in the cluster.

How do you remove/add members to the ZK member list? The issue you encounter 
might come from inconsistencies in ZK cluster.

Regards,

Thomas
________________________________
From: Charles-François Natali <cf.nat...@gmail.com>
Sent: Wednesday, 9 March 2022 23:44
To: user <user@mesos.apache.org>
Subject: [BULK]Re: Mesos and Zookeeper Dynamic Reconfiguration?

Hi Dan,

I don't think anyone has been looking at this, and i doubt we will, since we 
are quite low on resources.


Cheers,




On Tue, Mar 8, 2022, 19:01 Dan Leary 
<d...@touchplan.io<mailto:d...@touchplan.io>> wrote:
Been doing some testing with mesos 1.11.0 and zookeeper's "dynamic 
reconfiguration" capability.
https://zookeeper.apache.org/doc/r3.6.3/zookeeperReconfig.html<https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fzookeeper.apache.org%2Fdoc%2Fr3.6.3%2FzookeeperReconfig.html&data=04%7C01%7Ct.lange%40criteo.com%7C070fc8a5cd6341e9a91008da021e7876%7C2a35d8fd574d48e3927c8c398e225a01%7C1%7C0%7C637824627127329223%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=AjPUBCzFNsu0twnGptmACWDH8%2FTUktvVYKarGBRtGJI%3D&reserved=0>

Seems prima facie like mesos-master can handle a dynamic reconfig of the zk 
ensemble up to the point
where a new set of participants has been added to the ensemble and the old 
participants
have been demoted to non-voting followers.  But when the non-voting follower 
processes are
terminated the master logs seem to indicate that the masters keep trying and 
failing to reconnect
to the old zk leader, even though they've apparently received updates with the 
new ensemble participants.

Anybody have any insight into this?
Any plans to support zk dynamic reconfiguration in the future?
Seems like it could make for easier O/S maintenance of one's master/zk cluster 
hosts.

-Dan

Reply via email to