Mridul Muralidharan created SPARK-6165:
------------------------------------------
Summary: Aggregate and reduce should spool to disk and complete
Key: SPARK-6165
URL: https://issues.apache.org/jira/browse/SPARK-6165
Project: Spark
Issue Type: Improvement
Components: Spark Core
Affects Versions: 1.4.0
Reporter: Mridul Muralidharan
Priority: Minor
To prevent data from workers causing OOM at master, we have the property
'spark.driver.maxResultSize'.
But the OOM at master can be due to two reasons :
a) Data being sent from workers is too large - causing OOM at master.
b) Large number of moderate (to low) sized data being sent to master causing
OOM.
(For example: 500k tasks, 1k each)
spark.driver.maxResultSize protects against both - but (b) should be handled
more gracefully by master : example spool it to disk, aggregate without waiting
for entire result set to be fetched, etc.
Currently we are forced to use treeReduce and co to work around this problem :
adding to the latency of jobs.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
