[
https://issues.apache.org/jira/browse/SPARK-2530?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Andrew Or updated SPARK-2530:
-----------------------------
Affects Version/s: (was: 1.1.0)
1.0.1
> Relax incorrect assumption of one ExternalAppendOnlyMap per thread
> ------------------------------------------------------------------
>
> Key: SPARK-2530
> URL: https://issues.apache.org/jira/browse/SPARK-2530
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
> Affects Versions: 1.0.1
> Reporter: Andrew Or
>
> Originally reported by Matei.
> Our current implementation of EAOM assumes only one map is created per task.
> This is not true in the following case, however:
> {code}
> rdd1.join(rdd2).reduceByKey(...)
> {code}
> This is because reduce by key does a map side combine, which creates an EAOM
> that streams from an EAOM previously created by the same thread to aggregate
> values from the join.
> The more concerning thing is the following: we currently maintain a global
> shuffle memory map (thread ID -> memory used by that thread to shuffle). If
> we create two EAOMs in the same thread, the memory occupied by the first map
> may be clobbered by that occupied by the second. This has very adverse
> consequences if the first map is huge but the second is just starting out, in
> which case we end up believing that we use much less memory than we actually
> do.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
