[
https://issues.apache.org/jira/browse/ZOOKEEPER-1506?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14135537#comment-14135537
]
Flavio Junqueira commented on ZOOKEEPER-1506:
---------------------------------------------
I think there is a deeper problem here that I'm worried about. If you point a
name to a different server starting from scratch, then it is like the server
state has been wiped out from the perspective of the ZK replication, which can
lead to state loss in some corner cases. One way around this is to use
reconfiguration: remove and re-introduce the server. Is something like this
doable for you?
> Re-try DNS hostname -> IP resolution if node connection fails
> -------------------------------------------------------------
>
> Key: ZOOKEEPER-1506
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-1506
> Project: ZooKeeper
> Issue Type: Improvement
> Components: server
> Affects Versions: 3.4.5
> Environment: Ubuntu 11.04 64-bit
> Reporter: Mike Heffner
> Assignee: Michael Lasevich
> Priority: Critical
> Labels: patch
> Fix For: 3.5.1
>
> Attachments: ZOOKEEPER-1506.patch, ZOOKEEPER-1506.patch,
> ZOOKEEPER-1506.patch, zk-dns-caching-refresh.patch
>
>
> In our zoo.cfg we use hostnames to identify the ZK servers that are part of
> an ensemble. These hostnames are configured with a low (<= 60s) TTL and the
> IP address they map to can and does change. Our procedure for
> replacing/upgrading a ZK node is to boot an entirely new instance and remap
> the hostname to the new instance's IP address. Our expectation is that when
> the original ZK node is terminated/shutdown, the remaining nodes in the
> ensemble would reconnect to the new instance.
> However, what we are noticing is that the remaining ZK nodes do not attempt
> to re-resolve the hostname->IP mapping for the new server. Once the original
> ZK node is terminated, the existing servers continue to attempt contacting it
> at the old IP address. It would be great if the ZK servers could try to
> re-resolve the hostname when attempting to connect to a lost ZK server,
> instead of caching the lookup indefinitely. Currently we must do a rolling
> restart of the ZK ensemble after swapping a node -- which at three nodes
> means we periodically lose quorum.
> The exact method we are following is to boot new instances in EC2 and attach
> one, of a set of three, Elastic IP address. External to EC2 this IP address
> remains the same and maps to whatever instance it is attached to. Internal to
> EC2, the elastic IP hostname has a TTL of about 45-60 seconds and is remapped
> to the internal (10.x.y.z) address of the instance it is attached to.
> Therefore, in our case we would like ZK to pickup the new 10.x.y.z address
> that the elastic IP hostname gets mapped to and reconnect appropriately.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
