[
https://issues.apache.org/jira/browse/FLINK-2345?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14623806#comment-14623806
]
Stephan Ewen commented on FLINK-2345:
-------------------------------------
Can you give us a few more details about this effort?
I assume it only affects the "standalone" mode, as the YARN mode does not need
discovery (YARN manages nodes and containers). Also, in the high-availability
mode that is coming up, discovers of the masters/leader will happen with
zookeeper.
This proposal would hence affect the standalone mode in non-HA setups. That is
fair, such setups are common for experimenting.
Can you outline how the mechanism should work:
- How does it interoperate with the mode where people still configure the
cluster manually (I think that mode should keep existing)?
- What is the general mechanism? Multicast/broadcast messages?
- How do nodes know which master to discover, if multiple Flink setups run in
the same network?
BTW: Here are some notes about how discovery happens in the upcoming master
high-availability with zookeeper:
https://cwiki.apache.org/confluence/display/FLINK/JobManager+High+Availability
> Zero Configuration Cluster Setup
> --------------------------------
>
> Key: FLINK-2345
> URL: https://issues.apache.org/jira/browse/FLINK-2345
> Project: Flink
> Issue Type: New Feature
> Reporter: Suminda Dharmasena
>
> Best for of local clustering is using auto discovery. Also automatic tuning
> of cluster configuration parameters.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
