[
https://issues.apache.org/jira/browse/YARN-9571?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Craig Condit updated YARN-9571:
-------------------------------
Description:
YARN-4946 introduced logic which prevents applications from being removed from
the RM State Store unless log aggregation has reached a terminal state.
However, there are cases where log aggregation may not always complete or fail
cleanly (we've seen instances where it shows as NOT_STARTED on production
clusters), and in this case the state store will continue to build up large
numbers of applications.
We should make this behavior configurable. I propose to add a new configuration:
{code:java}
yarn.resourcemanager.completed-application-ttl-secs
{code}
This can be used to force removal of an application if it would otherwise be
purged from the state store if not for log aggregation status. If this
configuration is set to a positive value, it will be used in conjunction with
the application finish time to determine whether or not to purge the app. The
app would be removed if the max # of completed applications has been reached
AND either the app has completed more than ttl-secs seconds ago OR the app has
completed log aggregation. For backwards compatibility we can set the default
value of this parameter to zero, which would disable the new logic.
was:
YARN-4946 introduced logic which prevents applications from being removed from
the RM State Store unless log aggregation has reached a terminal state.
However, there are cases where log aggregation may not always complete or fail
cleanly (we've seen instances where it shows as NOT_STARTED on production
clusters), and in this case the state store will continue to build up large
numbers of applications.
We should make this behavior configurable. I propose to add a new configuration
(
yarn.resourcemanager.completed-application-ttl-secs) which can be used to force
removal of an application if it would otherwise be purged from the state store
if not for log aggregation status. If this configuration is set to a positive
value, it will be used in conjunction with the application finish time to
determine whether or not to purge the app. The app would be removed if the max
# of completed applications has been reached AND either the app has completed
more than ttl-secs seconds ago OR the app has completed log aggregation. For
backwards compatibility we can set the default value of this parameter to zero,
which would disable the new logic.
> RM state store purging should be configurable if log aggregation does not
> terminate
> -----------------------------------------------------------------------------------
>
> Key: YARN-9571
> URL: https://issues.apache.org/jira/browse/YARN-9571
> Project: Hadoop YARN
> Issue Type: Improvement
> Affects Versions: 3.2.0
> Reporter: Craig Condit
> Assignee: Craig Condit
> Priority: Major
>
> YARN-4946 introduced logic which prevents applications from being removed
> from the RM State Store unless log aggregation has reached a terminal state.
> However, there are cases where log aggregation may not always complete or
> fail cleanly (we've seen instances where it shows as NOT_STARTED on
> production clusters), and in this case the state store will continue to build
> up large numbers of applications.
> We should make this behavior configurable. I propose to add a new
> configuration:
> {code:java}
> yarn.resourcemanager.completed-application-ttl-secs
> {code}
> This can be used to force removal of an application if it would otherwise be
> purged from the state store if not for log aggregation status. If this
> configuration is set to a positive value, it will be used in conjunction with
> the application finish time to determine whether or not to purge the app. The
> app would be removed if the max # of completed applications has been reached
> AND either the app has completed more than ttl-secs seconds ago OR the app
> has completed log aggregation. For backwards compatibility we can set the
> default value of this parameter to zero, which would disable the new logic.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
