[
https://issues.apache.org/jira/browse/MESOS-2602?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14525250#comment-14525250
]
haosdent commented on MESOS-2602:
---------------------------------
How about save these status to zookeeper? And change the implementation of
/master/state.json to query data in zk? And for other service, they could
direct query zookeeper.
> Provide a way to "push" cluster state updates to a registered service.
> -----------------------------------------------------------------------
>
> Key: MESOS-2602
> URL: https://issues.apache.org/jira/browse/MESOS-2602
> Project: Mesos
> Issue Type: Improvement
> Reporter: Christos Kozyrakis
>
> At the moment, service discovery systems like Mesos-DNS must periodically
> pull the cluster state through state.json. This is extremely wasteful and
> will not scale to large clusters. If the polling interval is low, the master
> load will grow significantly. If the polling interval is high, there will be
> added latency to service discovery. Moreover, the way state.json works right
> now, one keeps reading the same information over and over again, including
> info about about tasks no longer running.
> We can design an endpoint that allows a "push" approach for state
> information. Here is one of the possible ways to set it up:
> - a service can hit the end point at (re)start to get information for all
> currently running tasks.
> - a service can also register itself to get receive updates to task state
> beyond that (ie, notifications of tasks starting/ending/etc). We may want to
> add some qualifiers here, since service discovery systems care only about
> certain types of updates.
> This can be implemented through direct messaging, through a message queue, by
> putting messages in etcd/zookeeper, etc. We should pick the way that is most
> scalable.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
