Hi Jungtaek Lim, SparkInterpreter uses scala REPL inside. Please see related issue https://issues.scala-lang.org/browse/SI-4331. There's a workaround in the description.
But I believe there will be no easy way to free up the memory completely, unless destroy and create scala REPL again. Thanks, moon On Thu, Dec 24, 2015 at 9:27 AM Jungtaek Lim <kabh...@gmail.com> wrote: > Forgot to add error log and stack traces, > > 15/12/16 11:17:02 INFO SchedulerFactory: Job > remoteInterpretJob_1450232220684 started by scheduler > org.apache.zeppelin.spark.SparkInterpreter2005736637 > 15/12/16 11:17:08 ERROR Job: Job failed > java.lang.OutOfMemoryError: Java heap space > at scala.reflect.internal.Names$class.enterChars(Names.scala:69) > at scala.reflect.internal.Names$class.newTermName(Names.scala:104) > at > scala.reflect.internal.SymbolTable.newTermName(SymbolTable.scala:13) > at scala.reflect.internal.Names$class.newTermName(Names.scala:113) > at > scala.reflect.internal.SymbolTable.newTermName(SymbolTable.scala:13) > at scala.reflect.internal.Names$class.newTypeName(Names.scala:116) > at > scala.reflect.internal.SymbolTable.newTypeName(SymbolTable.scala:13) > at scala.reflect.internal.Names$TypeName.newName(Names.scala:531) > at scala.reflect.internal.Names$TypeName.newName(Names.scala:513) > at scala.reflect.internal.Names$Name.append(Names.scala:424) > at > scala.reflect.internal.Symbols$Symbol.fullNameInternal(Symbols.scala:1044) > at > scala.reflect.internal.Symbols$Symbol.fullNameAsName(Symbols.scala:1047) > at > scala.reflect.internal.Symbols$Symbol.fullNameInternal(Symbols.scala:1044) > at > scala.reflect.internal.Symbols$Symbol.fullNameAsName(Symbols.scala:1047) > at > scala.reflect.internal.Symbols$Symbol.fullNameInternal(Symbols.scala:1044) > at > scala.reflect.internal.Symbols$Symbol.fullNameAsName(Symbols.scala:1047) > at > scala.reflect.internal.Symbols$Symbol.fullNameInternal(Symbols.scala:1044) > at > scala.reflect.internal.Symbols$Symbol.fullNameAsName(Symbols.scala:1047) > at > scala.reflect.internal.Symbols$Symbol.fullNameInternal(Symbols.scala:1044) > at > scala.reflect.internal.Symbols$Symbol.fullNameAsName(Symbols.scala:1047) > at > scala.reflect.internal.Symbols$Symbol.fullNameInternal(Symbols.scala:1044) > at > scala.reflect.internal.Symbols$Symbol.fullNameAsName(Symbols.scala:1047) > at > scala.reflect.internal.Symbols$Symbol.fullNameInternal(Symbols.scala:1044) > at > scala.reflect.internal.Symbols$Symbol.fullNameAsName(Symbols.scala:1047) > at > scala.reflect.internal.Symbols$Symbol.fullNameInternal(Symbols.scala:1044) > at > scala.reflect.internal.Symbols$Symbol.fullNameAsName(Symbols.scala:1047) > at > scala.reflect.internal.Symbols$Symbol.fullNameInternal(Symbols.scala:1044) > at > scala.reflect.internal.Symbols$Symbol.fullNameAsName(Symbols.scala:1047) > at > scala.reflect.internal.Symbols$Symbol.fullNameInternal(Symbols.scala:1044) > at > scala.reflect.internal.Symbols$Symbol.fullNameAsName(Symbols.scala:1047) > at > scala.reflect.internal.Symbols$Symbol.fullNameInternal(Symbols.scala:1044) > at > scala.reflect.internal.Symbols$Symbol.fullNameAsName(Symbols.scala:1047) > 15/12/16 11:17:08 INFO SchedulerFactory: Job > remoteInterpretJob_1450232220684 finished by scheduler > org.apache.zeppelin.spark.SparkInterpreter2005736637 > > Same logs are printed whenever run new job after OOME. > > > On Thu, Dec 24, 2015 at 9:25 AM, Jungtaek Lim <kabh...@gmail.com> wrote: > >> Hi users, >> >> I've met OOME when using spark interpreter and wish to resolve this issue. >> >> - Spark version: 1.4.1 + applying SPARK-11818 >> <http://issues.apache.org/jira/browse/SPARK-11818> >> - Spark cluster: Mesos 0.22.1 >> - Zeppelin: commit 1ba6e2a >> <https://github.com/apache/incubator-zeppelin/commit/1ba6e2a5969e475bc926943885c120f793266147> >> + >> applying ZEPPELIN-507 >> <https://issues.apache.org/jira/browse/ZEPPELIN-507> & ZEPPELIN-509 >> <https://issues.apache.org/jira/browse/ZEPPELIN-509> >> - loaded one fat driver jar via %dep >> >> I've run paragraph which dumps hbase table to hdfs several times, and >> takes memory histogram via "jmap -histo:live <pid>". >> Looking at histograms I can see that interpreter memory usages is >> increased whenever I run the paragraph. >> There could be spark app's memory leak, but nothing is clear so I'd like >> to find any other users who see the same behavior. >> >> Is anyone seeing same behavior, and could you share how to resolve? >> >> Thanks, >> Jungtaek Lim (HeartSaVioR) >> > > > > -- > Name : Jungtaek Lim > Blog : http://medium.com/@heartsavior > Twitter : http://twitter.com/heartsavior > LinkedIn : http://www.linkedin.com/in/heartsavior >
