[
https://issues.apache.org/jira/browse/FLINK-1442?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14292498#comment-14292498
]
ASF GitHub Bot commented on FLINK-1442:
---------------------------------------
Github user StephanEwen commented on the pull request:
https://github.com/apache/flink/pull/344#issuecomment-71548964
Looks good so far. I see that you removed the LRU code. Was that on purpose?
Leaving it in may be a good idea, because the soft references are cleared
in arbitrary order. It may make newer jobs disappear before older ones. Having
the LRU in would mean things behave as previously as long as the memory is
sufficient, and the soft reference clearing kicks in as a safety valve.
> Archived Execution Graph consumes too much memory
> -------------------------------------------------
>
> Key: FLINK-1442
> URL: https://issues.apache.org/jira/browse/FLINK-1442
> Project: Flink
> Issue Type: Bug
> Components: JobManager
> Affects Versions: 0.9
> Reporter: Stephan Ewen
> Assignee: Max Michels
>
> The JobManager archives the execution graphs, for analysis of jobs. The
> graphs may consume a lot of memory.
> Especially the execution edges in all2all connection patterns are extremely
> many and add up in memory consumption.
> The execution edges connect all parallel tasks. So for a all2all pattern
> between n and m tasks, there are n*m edges. For parallelism of multiple 100
> tasks, this can easily reach 100k objects and more, each with a set of
> metadata.
> I propose the following to solve that:
> 1. Clear all execution edges from the graph (majority of the memory
> consumers) when it is given to the archiver.
> 2. Have the map/list of the archived graphs behind a soft reference, to it
> will be removed under memory pressure before the JVM crashes. That may remove
> graphs from the history early, but is much preferable to the JVM crashing, in
> which case the graph is lost as well...
> 3. Long term: The graph should be archived somewhere else. Somthing like the
> History server used by Hadoop and Hive would be a good idea.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
