[ https://issues.apache.org/jira/browse/YARN-5685?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15530951#comment-15530951
]
Daniel Templeton edited comment on YARN-5685 at 9/28/16 9:37 PM:
-----------------------------------------------------------------
I think I finally get it. The net is that HA now requires either Curator or embedded to be
enabled. Letting the {{ZKRMStateStore}} handle the RM state transitions is no longer an option
and hasn't been for a while. There are only three ways {{transitionToActive()}} gets called:
# {{EmbeddedElectorService.becomeActive()}} -- embedded failover
# {{LeaderElectorService.isLeader()}} -- curator failover
# {{AdminService.transitionToActive()}} -- CLI
The question is what to do about non-embedded failover. I don't think it's worth making non-embedded
failover work again. I think we're better off declaring non-embedded failover to be dead
and removing the option altogether.
was (Author: templedf):
I think I finally get it. The net is that HA now requires either Curator or embedded to be
enabled. Letting the {{ZKRMStateStore}} handle the RM state transitions is no longer an option
and hasn't been for a while. There are only three ways {{transitionToActive()}} gets called:
# {{EmbeddedElectorService.becomeActive()}} -- embedded failover
# {{LeaderElectorService.isLeader()}} -- curator failover
# {{AdminService.transitionToActive()}} -- CLI
The question is what to do about non-embedded failover. I don't think it's worth making non-embedded
failover work again. I think we're better off declaring non-embedded failover to be dead
and removing the option altogether.
Given that this has been broken for a while and no one has noticed, I'm going to bump the
priority down a bit.
> Non-embedded HA failover is broken
> ----------------------------------
>
> Key: YARN-5685
> URL: https://issues.apache.org/jira/browse/YARN-5685
> Project: Hadoop YARN
> Issue Type: Bug
> Components: resourcemanager
> Affects Versions: 2.9.0, 3.0.0-alpha1
> Reporter: Daniel Templeton
> Assignee: Daniel Templeton
> Priority: Critical
>
> YARN-4559 broke RM HA when embedded automatic failover is disabled. The {{ZKRMStateStore}}
will now only start its monitoring thread when automatic failover not enabled (which is patently
useless). I presume the intended change was to have the monitoring thread started when automatic
failover is not *embedded*.
> If HA is enabled with automatic failover enabled and embedded failover disabled, all
RMs all come up in standby state. To make one of them active, the {{--forcemanual}} flag
must be used when manually triggering the state change. Should the active go down, the standby
will not become active and must be manually transitioned with the {{--forcemanual}} flag.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org
|