[SPARK-6847][CORE][STREAMING] Fix stack overflow issue when updateStateByKey is followed by a checkpointed dstream

Add a local property to indicate if checkpointing all RDDs that are marked with the checkpoint flag, and enable it in Streaming

Author: Shixiong Zhu <shixiong@databricks.com>

Closes #10934 from zsxwing/recursive-checkpoint.

3 files changed, 79 insertions, 2 deletions

diff --git a/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobGenerator.scala b/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobGenerator.scala
index a5a01e7763..a3ad5eaa40 100644
--- a/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobGenerator.scala
+++ b/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobGenerator.scala
@@ -20,6 +20,7 @@ package org.apache.spark.streaming.scheduler
 import scala.util.{Failure, Success, Try}
 
 import org.apache.spark.{Logging, SparkEnv}
+import org.apache.spark.rdd.RDD
 import org.apache.spark.streaming.{Checkpoint, CheckpointWriter, Time}
 import org.apache.spark.streaming.util.RecurringTimer
 import org.apache.spark.util.{Clock, EventLoop, ManualClock, Utils}
@@ -243,6 +244,10 @@ class JobGenerator(jobScheduler: JobScheduler) extends Logging {
     // Example: BlockRDDs are created in this thread, and it needs to access BlockManager
     // Update: This is probably redundant after threadlocal stuff in SparkEnv has been removed.
     SparkEnv.set(ssc.env)
+
+    // Checkpoint all RDDs marked for checkpointing to ensure their lineages are
+    // truncated periodically. Otherwise, we may run into stack overflows (SPARK-6847).
+    ssc.sparkContext.setLocalProperty(RDD.CHECKPOINT_ALL_MARKED_ANCESTORS, "true")
     Try {
       jobScheduler.receiverTracker.allocateBlocksToBatch(time) // allocate received blocks to batch
       graph.generateJobs(time) // generate jobs using allocated block
diff --git a/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobScheduler.scala b/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobScheduler.scala
index 9535c8e5b7..3fed3d8835 100644
--- a/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobScheduler.scala
+++ b/streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobScheduler.scala
@@ -23,10 +23,10 @@ import scala.collection.JavaConverters._
 import scala.util.Failure
 
 import org.apache.spark.Logging
-import org.apache.spark.rdd.PairRDDFunctions
+import org.apache.spark.rdd.{PairRDDFunctions, RDD}
 import org.apache.spark.streaming._
 import org.apache.spark.streaming.ui.UIUtils
-import org.apache.spark.util.{EventLoop, ThreadUtils, Utils}
+import org.apache.spark.util.{EventLoop, ThreadUtils}
 
 
 private[scheduler] sealed trait JobSchedulerEvent
@@ -210,6 +210,9 @@ class JobScheduler(val ssc: StreamingContext) extends Logging {
           s"""Streaming job from <a href="$batchUrl">$batchLinkText</a>""")
         ssc.sc.setLocalProperty(BATCH_TIME_PROPERTY_KEY, job.time.milliseconds.toString)
         ssc.sc.setLocalProperty(OUTPUT_OP_ID_PROPERTY_KEY, job.outputOpId.toString)
+        // Checkpoint all RDDs marked for checkpointing to ensure their lineages are
+        // truncated periodically. Otherwise, we may run into stack overflows (SPARK-6847).
+        ssc.sparkContext.setLocalProperty(RDD.CHECKPOINT_ALL_MARKED_ANCESTORS, "true")
 
         // We need to assign `eventLoop` to a temp variable. Otherwise, because
         // `JobScheduler.stop(false)` may set `eventLoop` to null when this method is running, then
diff --git a/streaming/src/test/scala/org/apache/spark/streaming/CheckpointSuite.scala b/streaming/src/test/scala/org/apache/spark/streaming/CheckpointSuite.scala
index 4a6b91fbc7..786703eb9a 100644
--- a/streaming/src/test/scala/org/apache/spark/streaming/CheckpointSuite.scala
+++ b/streaming/src/test/scala/org/apache/spark/streaming/CheckpointSuite.scala
@@ -821,6 +821,75 @@ class CheckpointSuite extends TestSuiteBase with DStreamCheckpointTester
     checkpointWriter.stop()
   }
 
+  test("SPARK-6847: stack overflow when updateStateByKey is followed by a checkpointed dstream") {
+    // In this test, there are two updateStateByKey operators. The RDD DAG is as follows:
+    //
+    //     batch 1            batch 2            batch 3     ...
+    //
+    // 1) input rdd          input rdd          input rdd
+    //       |                  |                  |
+    //       v                  v                  v
+    // 2) cogroup rdd   ---> cogroup rdd   ---> cogroup rdd  ...
+    //       |         /        |         /        |
+    //       v        /         v        /         v
+    // 3)  map rdd ---        map rdd ---        map rdd     ...
+    //       |                  |                  |
+    //       v                  v                  v
+    // 4) cogroup rdd   ---> cogroup rdd   ---> cogroup rdd  ...
+    //       |         /        |         /        |
+    //       v        /         v        /         v
+    // 5)  map rdd ---        map rdd ---        map rdd     ...
+    //
+    // Every batch depends on its previous batch, so "updateStateByKey" needs to do checkpoint to
+    // break the RDD chain. However, before SPARK-6847, when the state RDD (layer 5) of the second
+    // "updateStateByKey" does checkpoint, it won't checkpoint the state RDD (layer 3) of the first
+    // "updateStateByKey" (Note: "updateStateByKey" has already marked that its state RDD (layer 3)
+    // should be checkpointed). Hence, the connections between layer 2 and layer 3 won't be broken
+    // and the RDD chain will grow infinitely and cause StackOverflow.
+    //
+    // Therefore SPARK-6847 introduces "spark.checkpoint.checkpointAllMarked" to force checkpointing
+    // all marked RDDs in the DAG to resolve this issue. (For the previous example, it will break
+    // connections between layer 2 and layer 3)
+    ssc = new StreamingContext(master, framework, batchDuration)
+    val batchCounter = new BatchCounter(ssc)
+    ssc.checkpoint(checkpointDir)
+    val inputDStream = new CheckpointInputDStream(ssc)
+    val updateFunc = (values: Seq[Int], state: Option[Int]) => {
+      Some(values.sum + state.getOrElse(0))
+    }
+    @volatile var shouldCheckpointAllMarkedRDDs = false
+    @volatile var rddsCheckpointed = false
+    inputDStream.map(i => (i, i))
+      .updateStateByKey(updateFunc).checkpoint(batchDuration)
+      .updateStateByKey(updateFunc).checkpoint(batchDuration)
+      .foreachRDD { rdd =>
+        /**
+         * Find all RDDs that are marked for checkpointing in the specified RDD and its ancestors.
+         */
+        def findAllMarkedRDDs(rdd: RDD[_]): List[RDD[_]] = {
+          val markedRDDs = rdd.dependencies.flatMap(dep => findAllMarkedRDDs(dep.rdd)).toList
+          if (rdd.checkpointData.isDefined) {
+            rdd :: markedRDDs
+          } else {
+            markedRDDs
+          }
+        }
+
+        shouldCheckpointAllMarkedRDDs =
+          Option(rdd.sparkContext.getLocalProperty(RDD.CHECKPOINT_ALL_MARKED_ANCESTORS)).
+            map(_.toBoolean).getOrElse(false)
+
+        val stateRDDs = findAllMarkedRDDs(rdd)
+        rdd.count()
+        // Check the two state RDDs are both checkpointed
+        rddsCheckpointed = stateRDDs.size == 2 && stateRDDs.forall(_.isCheckpointed)
+      }
+    ssc.start()
+    batchCounter.waitUntilBatchesCompleted(1, 10000)
+    assert(shouldCheckpointAllMarkedRDDs === true)
+    assert(rddsCheckpointed === true)
+  }
+
   /**
    * Advances the manual clock on the streaming scheduler by given number of batches.
    * It also waits for the expected amount of time for each batch.