[SPARK-7766] KryoSerializerInstance reuse is unsafe when auto-reset is disabled

SPARK-3386 / apache#5606 modified the shuffle write path to re-use serializer instances across multiple calls to DiskBlockObjectWriter. It turns out that this introduced a very rare bug when using `KryoSerializer`: if auto-reset is disabled and reference-tracking is enabled, then we'll end up re-using the same serializer instance to write multiple output streams without calling `reset()` between write calls, which can lead to cases where objects in one file may contain references to objects that are in previous files, causing errors during deserialization.

This patch fixes this bug by calling `reset()` at the start of `serialize()` and `serializeStream()`. I also added a regression test which demonstrates that this problem only occurs when auto-reset is disabled and reference-tracking is enabled.

Author: Josh Rosen <[email protected]>

Closes apache#6293 from JoshRosen/kryo-instance-reuse-bug and squashes the following commits:

e19726d [Josh Rosen] Add fix for SPARK-7766.
71845e3 [Josh Rosen] Add failing regression test to trigger Kryo re-use bug

Author: JoshRosen

core/src/main/scala/org/apache/spark/serializer/KryoSerializer.scala

@@ -177,6 +177,7 @@ private[spark] class KryoSerializerInstance(ks: KryoSerializer) extends Serializ
 
   override def serialize[T: ClassTag](t: T): ByteBuffer = {
     output.clear()
+    kryo.reset() // We must reset in case this serializer instance was reused (see SPARK-7766)
     try {
       kryo.writeClassAndObject(output, t)
     } catch {
@@ -202,6 +203,7 @@ private[spark] class KryoSerializerInstance(ks: KryoSerializer) extends Serializ
   }
 
   override def serializeStream(s: OutputStream): SerializationStream = {
+    kryo.reset() // We must reset in case this serializer instance was reused (see SPARK-7766)
     new KryoSerializationStream(kryo, s)
   }
 

core/src/test/scala/org/apache/spark/serializer/KryoSerializerSuite.scala

@@ -17,6 +17,8 @@
 
 package org.apache.spark.serializer
 
+import java.io.ByteArrayOutputStream
+
 import scala.collection.mutable
 import scala.reflect.ClassTag
 
@@ -319,6 +321,37 @@ class KryoSerializerSuite extends FunSuite with SharedSparkContext {
     val ser2 = new KryoSerializer(conf).newInstance().asInstanceOf[KryoSerializerInstance]
     assert(!ser2.getAutoReset)
   }
+
+  private def testSerializerInstanceReuse(autoReset: Boolean, referenceTracking: Boolean): Unit = {
+    val conf = new SparkConf(loadDefaults = false)
+      .set("spark.kryo.referenceTracking", referenceTracking.toString)
+    if (!autoReset) {
+      conf.set("spark.kryo.registrator", classOf[RegistratorWithoutAutoReset].getName)
+    }
+    val ser = new KryoSerializer(conf)
+    val serInstance = ser.newInstance().asInstanceOf[KryoSerializerInstance]
+    assert (serInstance.getAutoReset() === autoReset)
+    val obj = ("Hello", "World")
+    def serializeObjects(): Array[Byte] = {
+      val baos = new ByteArrayOutputStream()
+      val serStream = serInstance.serializeStream(baos)
+      serStream.writeObject(obj)
+      serStream.writeObject(obj)
+      serStream.close()
+      baos.toByteArray
+    }
+    val output1: Array[Byte] = serializeObjects()
+    val output2: Array[Byte] = serializeObjects()
+    assert (output1 === output2)
+  }
+
+  // Regression test for SPARK-7766, an issue where disabling auto-reset and enabling
+  // reference-tracking would lead to corrupted output when serializer instances are re-used
+  for (referenceTracking <- Set(true, false); autoReset <- Set(true, false)) {
+    test(s"instance reuse with autoReset = $autoReset, referenceTracking = $referenceTracking") {
+      testSerializerInstanceReuse(autoReset = autoReset, referenceTracking = referenceTracking)
+    }
+  }
 }