hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Billie Rinaldi (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (YARN-6533) Race condition in writing service record to registry in yarn native services
Date Wed, 10 May 2017 14:33:04 GMT

    [ https://issues.apache.org/jira/browse/YARN-6533?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16004772#comment-16004772

Billie Rinaldi commented on YARN-6533:

I am not sure about the purpose of the YARN_ID. I saw in SelectByYarnPersistence and RMRegistryOperationsService
that sometimes it is used as a selection criterion for deleting service records. It looks
like the appId should not be encoded. I think the reason the containerId is encoded is that
it's used as a path in the registry, while the appId is not.

That said, I was worried that the YARN_ID was being used for deleting container service records,
but it does not look like that is true. It might only be used for deleting application service
records. Then my only concern would be the fact that the YARN_ID for the container record
was different in the initial registration service record vs. the update service record. This
would no longer be the case since we are removing the initial record. So maybe we could leave
the container YARN_ID unencoded? What do you think, [~jianhe]?

> Race condition in writing service record to registry in yarn native services
> ----------------------------------------------------------------------------
>                 Key: YARN-6533
>                 URL: https://issues.apache.org/jira/browse/YARN-6533
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>            Reporter: Billie Rinaldi
>            Assignee: Billie Rinaldi
>         Attachments: YARN-6533-yarn-native-services.001.patch, YARN-6533-yarn-native-services.002.patch
> The ServiceRecord is written twice, once when the container is initially registered and
again in the Docker provider once the IP has been obtained for the container. These occur
asynchronously, so the more important record (the one with the IP) can be overwritten by the
initial record. Only one record needs to be written, so we can stop writing the initial record
when the Docker provider is being used.

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