[PIG-4323] PackageConverter hanging in Spark

• Type: Sub-task
• Status: Patch Available
• Priority: Major
• Resolution: Unresolved
• Fix Version/s: spark-branch
• Component/s: spark
• Patch Info: Patch Available
• Assignee: Carlos Balduz
• Reporter: Carlos Balduz
• Created: 11/Nov/14 17:03
• Updated: 24/Feb/15 17:59

Description

After the patch introduced in PIG-4237, PackageConverter hangs for large jobs, since it has to deserialize JobConfiguration each time apply is called, making it way too slow:

public Tuple apply(final Tuple t) {
            initializeJobConf();
            ...

After having a job stuck for 30 minutes, I made a thread dump and saw that the main issue was:

Thread 11128: (state = IN_JAVA)
 - java.util.zip.InflaterInputStream.read(byte[], int, int) @bci=53, line=152 (Compiled frame; information may be imprecise)
 - java.util.zip.GZIPInputStream.read(byte[], int, int) @bci=17, line=116 (Compiled frame)
 - org.apache.hadoop.io.WritableUtils.readCompressedByteArray(java.io.DataInput) @bci=65, line=44 (Compiled frame)
 - org.apache.hadoop.io.WritableUtils.readCompressedString(java.io.DataInput) @bci=1, line=87 (Compiled frame)
 - org.apache.hadoop.io.WritableUtils.readCompressedStringArray(java.io.DataInput) @bci=29, line=185 (Compiled frame)
 - org.apache.hadoop.conf.Configuration.readFields(java.io.DataInput) @bci=37, line=2564 (Compiled frame)
 - org.apache.pig.backend.hadoop.executionengine.spark.KryoSerializer.deserializeJobConf(byte[]) @bci=24, line=80 (Compiled frame)
 - org.apache.pig.backend.hadoop.executionengine.spark.converter.PackageConverter$PackageFunction.initializeJobConf() @bci=4, line=60 (Compiled frame)
 - org.apache.pig.backend.hadoop.executionengine.spark.converter.PackageConverter$PackageFunction.apply(org.apache.pig.data.Tuple) @bci=1, line=67 (Compiled frame)
 - org.apache.pig.backend.hadoop.executionengine.spark.converter.PackageConverter$PackageFunction.apply(java.lang.Object) @bci=5, line=48 (Compiled frame)
 - scala.collection.Iterator$$anon$11.next() @bci=13, line=328 (Compiled frame)
 - scala.collection.convert.Wrappers$IteratorWrapper.next() @bci=4, line=30 (Compiled frame)
 - org.apache.pig.backend.hadoop.executionengine.spark.converter.POOutputConsumerIterator.readNext() @bci=44, line=61 (Compiled frame)
 - org.apache.pig.backend.hadoop.executionengine.spark.converter.POOutputConsumerIterator.readNext() @bci=162, line=78 (Compiled frame)
 - org.apache.pig.backend.hadoop.executionengine.spark.converter.POOutputConsumerIterator.hasNext() @bci=1, line=91 (Compiled frame)
 - org.apache.pig.backend.hadoop.executionengine.spark.converter.POOutputConsumerIterator.readNext() @bci=133, line=75 (Compiled frame)
 - org.apache.pig.backend.hadoop.executionengine.spark.converter.POOutputConsumerIterator.hasNext() @bci=1, line=91 (Compiled frame)
 - scala.collection.convert.Wrappers$JIteratorWrapper.hasNext() @bci=4, line=41 (Compiled frame)
 - scala.collection.Iterator$$anon$11.hasNext() @bci=4, line=327 (Compiled frame)
 - scala.collection.Iterator$$anon$11.hasNext() @bci=4, line=327 (Compiled frame)
 - scala.collection.Iterator$class.foreach(scala.collection.Iterator, scala.Function1) @bci=1, line=727 (Compiled frame)
 - scala.collection.AbstractIterator.foreach(scala.Function1) @bci=2, line=1157 (Interpreted frame)
 - org.apache.spark.shuffle.hash.HashShuffleWriter.write(scala.collection.Iterator) @bci=95, line=65 (Interpreted frame)
 - org.apache.spark.scheduler.ShuffleMapTask.runTask(org.apache.spark.TaskContext) @bci=170, line=68 (Interpreted frame)
 - org.apache.spark.scheduler.ShuffleMapTask.runTask(org.apache.spark.TaskContext) @bci=2, line=41 (Interpreted frame)
 - org.apache.spark.scheduler.Task.run(long) @bci=67, line=54 (Interpreted frame)
 - org.apache.spark.executor.Executor$TaskRunner.run() @bci=336, line=177 (Interpreted frame)
 - java.util.concurrent.ThreadPoolExecutor.runWorker(java.util.concurrent.ThreadPoolExecutor$Worker) @bci=95, line=1145 (Interpreted frame)
 - java.util.concurrent.ThreadPoolExecutor$Worker.run() @bci=5, line=615 (Interpreted frame)
 - java.lang.Thread.run() @bci=11, line=745 (Interpreted frame)

I have added a condition that checks whether it has already been initialized or not, making it work again.