hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Wilfred Spiegelenburg (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (YARN-7622) Allow fair-scheduler configuration on HDFS
Date Thu, 14 Dec 2017 23:33:00 GMT

    [ https://issues.apache.org/jira/browse/YARN-7622?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16291796#comment-16291796

Wilfred Spiegelenburg commented on YARN-7622:

Sorry for the delay in the review,

Using the AtomicLong is overkill in this case. You are only using the get and set on them
which negates the whole idea about the atomicity. We are not updating or incrementing the
value and the use of volatile will thus be enough, see the documentation of [atomic|https://docs.oracle.com/javase/8/docs/api/java/util/concurrent/atomic/package-summary.html]

I also looked a bit closer at the change and tried to see if current failures would still
fail and that we are just allowing hdfs as the storage. With the change that is made I can
currently get any configuration value to pass. We should not do this. It will then go wrong
when the serviceInit is called before we start the reload thread:
* Using a HTTP URL which passes the tests. This will then later cause failures while trying
to load the file since the filesystem is wrong.
* A bogus URL which would not work in the previous code setup but now passes and fails in
the service init with an exception.

Two things:
# limit the URL that can be set to filesystems we support loading the file from (HDFS, S3?)
# negative tests (config we don't allow should not work)

> Allow fair-scheduler configuration on HDFS
> ------------------------------------------
>                 Key: YARN-7622
>                 URL: https://issues.apache.org/jira/browse/YARN-7622
>             Project: Hadoop YARN
>          Issue Type: Improvement
>          Components: fairscheduler, resourcemanager
>            Reporter: Greg Phillips
>            Assignee: Greg Phillips
>            Priority: Minor
>         Attachments: YARN-7622.001.patch, YARN-7622.002.patch, YARN-7622.003.patch
> The FairScheduler requires the allocation file to be hosted on the local filesystem on
the RM node(s). Allowing HDFS to store the allocation file will provide improved redundancy,
more options for scheduler updates, and RM failover consistency in HA.

This message was sent by Atlassian JIRA

To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org

View raw message