[
https://issues.apache.org/jira/browse/YARN-5375?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15435429#comment-15435429
]
Varun Saxena commented on YARN-5375:
------------------------------------
There are many events where draining of both rm dispatcher and state store
dispatcher is not required and as Rohith pointed out in some places, we may
need to drain RM Dispatcher after state store dispatcher drain as well. There
could be different sequences as well. Should we create 2 drain methods and call
each one as per need ?
Also many of the waitForState can be eliminated by calling a sequence of
drains. For different kind of events, we may have to decide what do we do.
Maybe for some regularly checked states, we can add some specific methods in
MockRM in addition to generic waitForState methods with customary draining of
RM dispatcher and state store dispatcher and again RM dispatcher. This as you
said can reduce test times.
Thoughts ?
For instance, after Node HB we typically check if an attempt has transferred to
ALLOCATED state in many of the tests. Here, first rm dispatcher must be drained
to process node update to be sent to Scheduler. As with the changes in patch,
scheduler events will be processed immediately, it will lead to App Attempt
event for Container being allocated (from RM Container). This would again
require draining of RM dispatcher queue which would lead to a state store event
and from State store we will again send a ATTEMPT_SAVED event back to RM App
attempt, which would again require draining of RM dispatcher queue. And only
then will attempt state change to ALLOCATED.
> invoke MockRM#drainEvents implicitly in MockRM methods to reduce test failures
> ------------------------------------------------------------------------------
>
> Key: YARN-5375
> URL: https://issues.apache.org/jira/browse/YARN-5375
> Project: Hadoop YARN
> Issue Type: Improvement
> Reporter: sandflee
> Assignee: sandflee
> Attachments: YARN-5375.01.patch, YARN-5375.03.patch,
> YARN-5375.04.patch, YARN-5375.05.patch, YARN-5375.06.patch
>
>
> seen many test failures related to RMApp/RMAppattempt comes to some state but
> some event are not processed in rm event queue or scheduler event queue,
> cause test failure, seems we could implicitly invokes drainEvents(should also
> drain sheduler event) in some mockRM method like waitForState
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
