[SPARK-16550][SPARK-17042][CORE] Certain classes fail to deserialize in block manager replication

## What changes were proposed in this pull request?

This is a straightforward clone of JoshRosen 's original patch. I have follow-up changes to fix block replication for repl-defined classes as well, but those appear to be flaking tests so I'm going to leave that for SPARK-17042

## How was this patch tested?

End-to-end test in ReplSuite (also more tests in DistributedSuite from the original patch).

Author: Eric Liang <ekl@databricks.com>

Closes #14311 from ericl/spark-16550.

3 files changed, 46 insertions, 58 deletions

diff --git a/core/src/main/scala/org/apache/spark/serializer/SerializerManager.scala b/core/src/main/scala/org/apache/spark/serializer/SerializerManager.scala
index 9dc274c9fe..07caadbe40 100644
--- a/core/src/main/scala/org/apache/spark/serializer/SerializerManager.scala
+++ b/core/src/main/scala/org/apache/spark/serializer/SerializerManager.scala
@@ -68,7 +68,7 @@ private[spark] class SerializerManager(defaultSerializer: Serializer, conf: Spar
    * loaded yet. */
   private lazy val compressionCodec: CompressionCodec = CompressionCodec.createCodec(conf)
 
-  private def canUseKryo(ct: ClassTag[_]): Boolean = {
+  def canUseKryo(ct: ClassTag[_]): Boolean = {
     primitiveAndPrimitiveArrayClassTags.contains(ct) || ct == stringClassTag
   }
 
@@ -128,8 +128,18 @@ private[spark] class SerializerManager(defaultSerializer: Serializer, conf: Spar
 
   /** Serializes into a chunked byte buffer. */
   def dataSerialize[T: ClassTag](blockId: BlockId, values: Iterator[T]): ChunkedByteBuffer = {
+    dataSerializeWithExplicitClassTag(blockId, values, implicitly[ClassTag[T]])
+  }
+
+  /** Serializes into a chunked byte buffer. */
+  def dataSerializeWithExplicitClassTag(
+      blockId: BlockId,
+      values: Iterator[_],
+      classTag: ClassTag[_]): ChunkedByteBuffer = {
     val bbos = new ChunkedByteBufferOutputStream(1024 * 1024 * 4, ByteBuffer.allocate)
-    dataSerializeStream(blockId, bbos, values)
+    val byteStream = new BufferedOutputStream(bbos)
+    val ser = getSerializer(classTag).newInstance()
+    ser.serializeStream(wrapForCompression(blockId, byteStream)).writeAll(values).close()
     bbos.toChunkedByteBuffer
   }
 
diff --git a/core/src/main/scala/org/apache/spark/storage/BlockManager.scala b/core/src/main/scala/org/apache/spark/storage/BlockManager.scala
index 015e71d126..fe84652798 100644
--- a/core/src/main/scala/org/apache/spark/storage/BlockManager.scala
+++ b/core/src/main/scala/org/apache/spark/storage/BlockManager.scala
@@ -498,7 +498,8 @@ private[spark] class BlockManager(
         diskStore.getBytes(blockId)
       } else if (level.useMemory && memoryStore.contains(blockId)) {
         // The block was not found on disk, so serialize an in-memory copy:
-        serializerManager.dataSerialize(blockId, memoryStore.getValues(blockId).get)
+        serializerManager.dataSerializeWithExplicitClassTag(
+          blockId, memoryStore.getValues(blockId).get, info.classTag)
       } else {
         handleLocalReadFailure(blockId)
       }
@@ -973,8 +974,16 @@ private[spark] class BlockManager(
         if (level.replication > 1) {
           val remoteStartTime = System.currentTimeMillis
           val bytesToReplicate = doGetLocalBytes(blockId, info)
+          // [SPARK-16550] Erase the typed classTag when using default serialization, since
+          // NettyBlockRpcServer crashes when deserializing repl-defined classes.
+          // TODO(ekl) remove this once the classloader issue on the remote end is fixed.
+          val remoteClassTag = if (!serializerManager.canUseKryo(classTag)) {
+            scala.reflect.classTag[Any]
+          } else {
+            classTag
+          }
           try {
-            replicate(blockId, bytesToReplicate, level, classTag)
+            replicate(blockId, bytesToReplicate, level, remoteClassTag)
           } finally {
             bytesToReplicate.dispose()
           }
diff --git a/core/src/test/scala/org/apache/spark/DistributedSuite.scala b/core/src/test/scala/org/apache/spark/DistributedSuite.scala
index 6beae842b0..4ee0e00fde 100644
--- a/core/src/test/scala/org/apache/spark/DistributedSuite.scala
+++ b/core/src/test/scala/org/apache/spark/DistributedSuite.scala
@@ -149,61 +149,16 @@ class DistributedSuite extends SparkFunSuite with Matchers with LocalSparkContex
     sc.parallelize(1 to 10).count()
   }
 
-  test("caching") {
+  private def testCaching(storageLevel: StorageLevel): Unit = {
     sc = new SparkContext(clusterUrl, "test")
-    val data = sc.parallelize(1 to 1000, 10).cache()
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-  }
-
-  test("caching on disk") {
-    sc = new SparkContext(clusterUrl, "test")
-    val data = sc.parallelize(1 to 1000, 10).persist(StorageLevel.DISK_ONLY)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-  }
-
-  test("caching in memory, replicated") {
-    sc = new SparkContext(clusterUrl, "test")
-    val data = sc.parallelize(1 to 1000, 10).persist(StorageLevel.MEMORY_ONLY_2)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-  }
-
-  test("caching in memory, serialized, replicated") {
-    sc = new SparkContext(clusterUrl, "test")
-    val data = sc.parallelize(1 to 1000, 10).persist(StorageLevel.MEMORY_ONLY_SER_2)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-  }
-
-  test("caching on disk, replicated") {
-    sc = new SparkContext(clusterUrl, "test")
-    val data = sc.parallelize(1 to 1000, 10).persist(StorageLevel.DISK_ONLY_2)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-  }
-
-  test("caching in memory and disk, replicated") {
-    sc = new SparkContext(clusterUrl, "test")
-    val data = sc.parallelize(1 to 1000, 10).persist(StorageLevel.MEMORY_AND_DISK_2)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-  }
-
-  test("caching in memory and disk, serialized, replicated") {
-    sc = new SparkContext(clusterUrl, "test")
-    val data = sc.parallelize(1 to 1000, 10).persist(StorageLevel.MEMORY_AND_DISK_SER_2)
-
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
-    assert(data.count() === 1000)
+    sc.jobProgressListener.waitUntilExecutorsUp(2, 30000)
+    val data = sc.parallelize(1 to 1000, 10)
+    val cachedData = data.persist(storageLevel)
+    assert(cachedData.count === 1000)
+    assert(sc.getExecutorStorageStatus.map(_.rddBlocksById(cachedData.id).size).sum ===
+      storageLevel.replication * data.getNumPartitions)
+    assert(cachedData.count === 1000)
+    assert(cachedData.count === 1000)
 
     // Get all the locations of the first partition and try to fetch the partitions
     // from those locations.
@@ -221,6 +176,20 @@ class DistributedSuite extends SparkFunSuite with Matchers with LocalSparkContex
     }
   }
 
+  Seq(
+    "caching" -> StorageLevel.MEMORY_ONLY,
+    "caching on disk" -> StorageLevel.DISK_ONLY,
+    "caching in memory, replicated" -> StorageLevel.MEMORY_ONLY_2,
+    "caching in memory, serialized, replicated" -> StorageLevel.MEMORY_ONLY_SER_2,
+    "caching on disk, replicated" -> StorageLevel.DISK_ONLY_2,
+    "caching in memory and disk, replicated" -> StorageLevel.MEMORY_AND_DISK_2,
+    "caching in memory and disk, serialized, replicated" -> StorageLevel.MEMORY_AND_DISK_SER_2
+  ).foreach { case (testName, storageLevel) =>
+    test(testName) {
+      testCaching(storageLevel)
+    }
+  }
+
   test("compute without caching when no partitions fit in memory") {
     val size = 10000
     val conf = new SparkConf()