[
https://issues.apache.org/jira/browse/YARN-4983?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15256481#comment-15256481
]
Junping Du commented on YARN-4983:
----------------------------------
Thanks [~gtCarrera9] for reporting the issue and delivering the fix.
The patch looks good in overall. There are some questions and comments below.
I think our work here is trying to differentiate the metric
initialization/re-initialization with RM first time start with RM transition
from standby. Isn't it? If so, below code:
{noformat}
@@ -1114,7 +1127,7 @@ void reinitialize(boolean initialize) {
QueueMetrics.clearQueueMetrics();
if (initialize) {
resetDispatcher();
- createAndInitActiveServices();
+ createAndInitActiveServices(true);
}
}
{noformat}
Shall we reuse {{boolean initialize}} instead of set true directly? I saw there
are two caller path: one is from RM.serviceStart() and the other is from
transition to standby from active, marked by boolean value of initialize. I
think we should also differentiate the situation here?
In addition, I think our RM start logic is a bit redundant when RM HA is
enabled: we always do serviceInit() first, then serviceStart(). However, within
serviceStart(), we always put RM to standby mode first and do reinitialize
again. Any special reason that we cannot initialize once to be in standby mode?
CC [~jianhe], [~ka...@cloudera.com].
> JVM and UGI metrics disappear after RM is once transitioned to standby mode
> ---------------------------------------------------------------------------
>
> Key: YARN-4983
> URL: https://issues.apache.org/jira/browse/YARN-4983
> Project: Hadoop YARN
> Issue Type: Bug
> Reporter: Li Lu
> Assignee: Li Lu
> Attachments: YARN-4983-trunk.000.patch, YARN-4983-trunk.001.patch,
> YARN-4983-trunk.002.patch
>
>
> When get transitioned to standby, the RM will shutdown the existing metric
> system and relaunch a new one. This will cause the jvm metrics and ugi
> metrics to miss in the new metric system.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
