[SPARK-16550] Caching data with replication doesn't replicate data

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.0.0
• Fix Version/s: 2.0.1, 2.1.0
• Component/s: Block Manager, Spark Core
• Assignee: Eric Liang
• Reporter: Shubham Chopra
• Created: 14/Jul/16 19:05
• Updated: 22/Aug/16 23:32
• Resolved: 22/Aug/16 23:31

Description

Caching multiple replicas of blocks is currently broken. The following examples show replication doesn't happen for various use-cases:

These were run using Spark 2.0.0-preview, in local-cluster[2,1,1024] mode

case class TestInteger(i: Int)
val data = sc.parallelize((1 to 1000).map(TestInteger(_)), 10).persist(MEMORY_ONLY_2)
data.count

sc.getExecutorStorageStatus.map(s => s.rddBlocksById(data.id).size).sum shows only 10 blocks as opposed to the expected 20
Block replication fails on the executors with a java.lang.RuntimeException: java.lang.ClassNotFoundException: $line14.$read$$iw$$iw$TestInteger

val data1 = sc.parallelize(1 to 1000, 10).persist(org.apache.spark.storage.StorageLevel.MEMORY_ONLY_2)
data1.count

Block replication again fails with the following errors:
16/07/14 14:50:40 ERROR TransportRequestHandler: Error while invoking RpcHandler#receive() on RPC id 8567643992794608648
com.esotericsoftware.kryo.KryoException: Encountered unregistered class ID: 13994
	at com.esotericsoftware.kryo.util.DefaultClassResolver.readClass(DefaultClassResolver.java:137)
	at com.esotericsoftware.kryo.Kryo.readClass(Kryo.java:670)
	at com.esotericsoftware.kryo.Kryo.readClassAndObject(Kryo.java:781)
	at org.apache.spark.serializer.KryoDeserializationStream.readObject(KryoSerializer.scala:229)
	at org.apache.spark.serializer.DeserializationStream$$anon$1.getNext(Serializer.scala:169)
	at org.apache.spark.util.NextIterator.hasNext(NextIterator.scala:73)
	at org.apache.spark.storage.memory.MemoryStore.putIteratorAsValues(MemoryStore.scala:213)
	at org.apache.spark.storage.BlockManager$$anonfun$doPutBytes$1.apply(BlockManager.scala:775)
	at org.apache.spark.storage.BlockManager$$anonfun$doPutBytes$1.apply(BlockManager.scala:753)

sc.getExecutorStorageStatus.map(s => s.rddBlocksById(data1.id).size).sum again shows 10 blocks

Caching serialized data works for native types, but not for custom classes

val data3 = sc.parallelize(1 to 1000, 10).persist(MEMORY_ONLY_SER_2)
data3.count

works as intended.

But

val data4 = sc.parallelize((1 to 1000).map(TestInteger(_)), 10).persist(MEMORY_ONLY_SER_2)
data4.count

Again doesn't replicate data and executors show the same ClassNotFoundException

These examples worked fine and showed expected results with Spark 1.6.2

Issue Links

• links to: [Github] Pull Request #14311 (ericl)