[
https://issues.apache.org/jira/browse/YARN-7790?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Wangda Tan updated YARN-7790:
-----------------------------
Description:
This is not a new issue but async scheduling makes it worse:
In sync scheduling, if an AM container allocated to a node, it assumes node
just heartbeat to RM, and AM launcher will connect NM to launch the container.
Even though it is possible that NM crashes after the heartbeat, which causes AM
hangs for a while. But it is related rare.
In async scheduling world, multiple AM containers can be placed on a
problematic NM, which could cause application hangs easily. Discussed with
[~sunilg] and [~jianhe] , we need one fix:
When async scheduling enabled:
1) Skip node which missed X node heartbeat.
And in addition, it's better to reduce wait time by setting following configs
to earlier fail a container being launched on a NM with connection issue.
{code:java}
RetryPolicy retryPolicy =
createRetryPolicy(conf,
YarnConfiguration.CLIENT_NM_CONNECT_MAX_WAIT_MS,
YarnConfiguration.DEFAULT_CLIENT_NM_CONNECT_MAX_WAIT_MS,
YarnConfiguration.CLIENT_NM_CONNECT_RETRY_INTERVAL_MS,
YarnConfiguration.DEFAULT_CLIENT_NM_CONNECT_RETRY_INTERVAL_MS);
{code}
was:
This is not a new issue but async scheduling makes it worse:
In sync scheduling, if an AM container allocated to a node, it assumes node
just heartbeat to RM, and in the same response, it will be sent back to NM.
Even though it is possible that NM crashes after the heartbeat, which causes AM
hangs for 10 mins. But it is related rare.
In async scheduling world, multiple AM containers can be placed on a
problematic NM, which could cause application hangs for long time. Discussed
with [~sunilg] , we need at least two fixes:
When async scheduling enabled:
1) Skip node which missed X node heartbeat.
2) Kill AM container in ALLOCATED state on a node which missed Y node heartbeat.
Credit to [[email protected]] who found the issue.
> Improve Capacity Scheduler Async Scheduling to better handle node failures
> --------------------------------------------------------------------------
>
> Key: YARN-7790
> URL: https://issues.apache.org/jira/browse/YARN-7790
> Project: Hadoop YARN
> Issue Type: Bug
> Reporter: Sumana Sathish
> Assignee: Wangda Tan
> Priority: Critical
> Attachments: YARN-7790.001.patch
>
>
> This is not a new issue but async scheduling makes it worse:
> In sync scheduling, if an AM container allocated to a node, it assumes node
> just heartbeat to RM, and AM launcher will connect NM to launch the
> container. Even though it is possible that NM crashes after the heartbeat,
> which causes AM hangs for a while. But it is related rare.
> In async scheduling world, multiple AM containers can be placed on a
> problematic NM, which could cause application hangs easily. Discussed with
> [~sunilg] and [~jianhe] , we need one fix:
> When async scheduling enabled:
> 1) Skip node which missed X node heartbeat.
> And in addition, it's better to reduce wait time by setting following configs
> to earlier fail a container being launched on a NM with connection issue.
> {code:java}
> RetryPolicy retryPolicy =
> createRetryPolicy(conf,
> YarnConfiguration.CLIENT_NM_CONNECT_MAX_WAIT_MS,
> YarnConfiguration.DEFAULT_CLIENT_NM_CONNECT_MAX_WAIT_MS,
> YarnConfiguration.CLIENT_NM_CONNECT_RETRY_INTERVAL_MS,
> YarnConfiguration.DEFAULT_CLIENT_NM_CONNECT_RETRY_INTERVAL_MS);
> {code}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
