hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jason Lowe (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (YARN-8385) Clean local directories when a container is killed
Date Thu, 05 Jul 2018 19:10:00 GMT

    [ https://issues.apache.org/jira/browse/YARN-8385?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16534030#comment-16534030
] 

Jason Lowe commented on YARN-8385:
----------------------------------

The container's working directories are cleaned up by YARN immediately after the container
terminates (unless the debug delay is configured on the NM).  The application directories
are not cleaned up until the application completes, as that is where a container can leave
local data that may be accessed by a subsequent container on the node or via an auxiliary
service (e.g.: shuffle data served up by a shuffle handler auxiliary service as is done for
MapReduce, Tez, and Spark).

Are you sure the data is being placed in the container's working directory and not the application
directory?

> Clean local directories when a container is killed
> --------------------------------------------------
>
>                 Key: YARN-8385
>                 URL: https://issues.apache.org/jira/browse/YARN-8385
>             Project: Hadoop YARN
>          Issue Type: Bug
>    Affects Versions: 2.7.0
>            Reporter: Marco Gaido
>            Priority: Major
>
> In long running applications, it may happen that many containers are created and killed.
A use case is Spark Thrift Server when dynamic allocation is enabled. A lot of containers
are killed and the application keeps running indefinitely.
> Currently, YARN seems to remove the local directories only when the whole application
terminates. In the scenario described above, this can cause serious resource leakages. Please,
check https://issues.apache.org/jira/browse/SPARK-22575.
> I think YARN should clean up all the local directories of a container when it is killed
and not when the whole application terminates.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

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


Mime
View raw message