[ https://issues.apache.org/jira/browse/BEAM-7647?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Etienne Chauchot updated BEAM-7647: ----------------------------------- Description: In combine globally: {code:java} Iterable<WindowedValue<OutputT>> output = sparkCombineFn.extractOutput(maybeAccumulated.get()); outRdd = context .getSparkContext() .parallelize(CoderHelpers.toByteArrays(output, wvoCoder)) .map(CoderHelpers.fromByteFunction(wvoCoder)); {code} => risk of OOM in the list, get data to a single worker (the driver) was: In combine globally: {code:java} Iterable<WindowedValue<OutputT>> output = sparkCombineFn.extractOutput(maybeAccumulated.get()); outRdd = context .getSparkContext() .parallelize(CoderHelpers.toByteArrays(output, wvoCoder)) .map(CoderHelpers.fromByteFunction(wvoCoder)); {code} => risk of OOM in the list, shuffle data to a single worker (the driver) > CombineGlobally translation is risky and not very performant. > ------------------------------------------------------------- > > Key: BEAM-7647 > URL: https://issues.apache.org/jira/browse/BEAM-7647 > Project: Beam > Issue Type: Improvement > Components: runner-spark > Reporter: Etienne Chauchot > Priority: Major > > In combine globally: > {code:java} > Iterable<WindowedValue<OutputT>> output = > sparkCombineFn.extractOutput(maybeAccumulated.get()); > outRdd = > context > .getSparkContext() > .parallelize(CoderHelpers.toByteArrays(output, wvoCoder)) > .map(CoderHelpers.fromByteFunction(wvoCoder)); > {code} > => risk of OOM in the list, get data to a single worker (the driver) -- This message was sent by Atlassian JIRA (v7.6.3#76005)