[SPARK-4835] Disable validateOutputSpecs for Spark Streaming jobs

This patch disables output spec. validation for jobs launched through Spark Streaming, since this interferes with checkpoint recovery.

Hadoop OutputFormats have a `checkOutputSpecs` method which performs certain checks prior to writing output, such as checking whether the output directory already exists.  SPARK-1100 added checks for FileOutputFormat, SPARK-1677 (#947) added a SparkConf configuration to disable these checks, and SPARK-2309 (#1088) extended these checks to run for all OutputFormats, not just FileOutputFormat.

In Spark Streaming, we might have to re-process a batch during checkpoint recovery, so `save` actions may be called multiple times.  In addition to `DStream`'s own save actions, users might use `transform` or `foreachRDD` and call the `RDD` and `PairRDD` save actions.  When output spec. validation is enabled, the second calls to these actions will fail due to existing output.

This patch automatically disables output spec. validation for jobs submitted by the Spark Streaming scheduler.  This is done by using Scala's `DynamicVariable` to propagate the bypass setting without having to mutate SparkConf or introduce a global variable.

Author: Josh Rosen <[email protected]>

Closes #3832 from JoshRosen/SPARK-4835 and squashes the following commits:

36eaf35 [Josh Rosen] Add comment explaining use of transform() in test.
6485cf8 [Josh Rosen] Add test case in Streaming; fix bug for transform()
7b3e06a [Josh Rosen] Remove Streaming-specific setting to undo this change; update conf. guide
bf9094d [Josh Rosen] Revise disableOutputSpecValidation() comment to not refer to Spark Streaming.
e581d17 [Josh Rosen] Deduplicate isOutputSpecValidationEnabled logic.
762e473 [Josh Rosen] [SPARK-4835] Disable validateOutputSpecs for Spark Streaming jobs.

(cherry picked from commit 939ba1f)
Signed-off-by: Tathagata Das <[email protected]>

Author: JoshRosen

core/src/main/scala/org/apache/spark/rdd/PairRDDFunctions.scala

@@ -25,6 +25,7 @@ import scala.collection.{Map, mutable}
 import scala.collection.JavaConversions._
 import scala.collection.mutable.ArrayBuffer
 import scala.reflect.ClassTag
+import scala.util.DynamicVariable
 
 import com.clearspring.analytics.stream.cardinality.HyperLogLogPlus
 import org.apache.hadoop.conf.{Configurable, Configuration}
@@ -960,7 +961,7 @@ class PairRDDFunctions[K, V](self: RDD[(K, V)])
     val outfmt = job.getOutputFormatClass
     val jobFormat = outfmt.newInstance
 
-    if (self.conf.getBoolean("spark.hadoop.validateOutputSpecs", true)) {
+    if (isOutputSpecValidationEnabled) {
       // FileOutputFormat ignores the filesystem parameter
       jobFormat.checkOutputSpecs(job)
     }
@@ -1038,7 +1039,7 @@ class PairRDDFunctions[K, V](self: RDD[(K, V)])
     logDebug("Saving as hadoop file of type (" + keyClass.getSimpleName + ", " +
       valueClass.getSimpleName + ")")
 
-    if (self.conf.getBoolean("spark.hadoop.validateOutputSpecs", true)) {
+    if (isOutputSpecValidationEnabled) {
       // FileOutputFormat ignores the filesystem parameter
       val ignoredFs = FileSystem.get(hadoopConf)
       hadoopConf.getOutputFormat.checkOutputSpecs(ignoredFs, hadoopConf)
@@ -1113,8 +1114,22 @@ class PairRDDFunctions[K, V](self: RDD[(K, V)])
   private[spark] def valueClass: Class[_] = vt.runtimeClass
 
   private[spark] def keyOrdering: Option[Ordering[K]] = Option(ord)
+
+  // Note: this needs to be a function instead of a 'val' so that the disableOutputSpecValidation
+  // setting can take effect:
+  private def isOutputSpecValidationEnabled: Boolean = {
+    val validationDisabled = PairRDDFunctions.disableOutputSpecValidation.value
+    val enabledInConf = self.conf.getBoolean("spark.hadoop.validateOutputSpecs", true)
+    enabledInConf && !validationDisabled
+  }
 }
 
 private[spark] object PairRDDFunctions {
   val RECORDS_BETWEEN_BYTES_WRITTEN_METRIC_UPDATES = 256
+
+  /**
+   * Allows for the `spark.hadoop.validateOutputSpecs` checks to be disabled on a case-by-case
+   * basis; see SPARK-4835 for more details.
+   */
+  val disableOutputSpecValidation: DynamicVariable[Boolean] = new DynamicVariable[Boolean](false)
 }

docs/configuration.md

@@ -709,7 +709,9 @@ Apart from these, the following properties are also available, and may be useful
     <td>If set to true, validates the output specification (e.g. checking if the output directory already exists)
     used in saveAsHadoopFile and other variants. This can be disabled to silence exceptions due to pre-existing
     output directories. We recommend that users do not disable this except if trying to achieve compatibility with
-    previous versions of Spark. Simply use Hadoop's FileSystem API to delete output directories by hand.</td>
+    previous versions of Spark. Simply use Hadoop's FileSystem API to delete output directories by hand.
+    This setting is ignored for jobs generated through Spark Streaming's StreamingContext, since
+    data may need to be rewritten to pre-existing output directories during checkpoint recovery.</td>
 </tr>
 <tr>
     <td><code>spark.hadoop.cloneConf</code></td>

streaming/src/main/scala/org/apache/spark/streaming/dstream/DStream.scala

@@ -26,7 +26,7 @@ import scala.reflect.ClassTag
 import scala.util.matching.Regex
 
 import org.apache.spark.{Logging, SparkException}
-import org.apache.spark.rdd.{BlockRDD, RDD}
+import org.apache.spark.rdd.{BlockRDD, PairRDDFunctions, RDD}
 import org.apache.spark.storage.StorageLevel
 import org.apache.spark.streaming._
 import org.apache.spark.streaming.StreamingContext._
@@ -293,7 +293,13 @@ abstract class DStream[T: ClassTag] (
         // set this DStream's creation site, generate RDDs and then restore the previous call site.
         val prevCallSite = ssc.sparkContext.getCallSite()
         ssc.sparkContext.setCallSite(creationSite)
-        val rddOption = compute(time)
+        // Disable checks for existing output directories in jobs launched by the streaming
+        // scheduler, since we may need to write output to an existing directory during checkpoint
+        // recovery; see SPARK-4835 for more details. We need to have this call here because
+        // compute() might cause Spark jobs to be launched.
+        val rddOption = PairRDDFunctions.disableOutputSpecValidation.withValue(true) {
+          compute(time)
+        }
         ssc.sparkContext.setCallSite(prevCallSite)
 
         rddOption.foreach { case newRDD =>

streaming/src/main/scala/org/apache/spark/streaming/dstream/TransformedDStream.scala

@@ -17,7 +17,7 @@
 
 package org.apache.spark.streaming.dstream
 
-import org.apache.spark.rdd.RDD
+import org.apache.spark.rdd.{PairRDDFunctions, RDD}
 import org.apache.spark.streaming.{Duration, Time}
 import scala.reflect.ClassTag
 

streaming/src/main/scala/org/apache/spark/streaming/scheduler/JobScheduler.scala

@@ -22,6 +22,7 @@ import scala.collection.JavaConversions._
 import java.util.concurrent.{TimeUnit, ConcurrentHashMap, Executors}
 import akka.actor.{ActorRef, Actor, Props}
 import org.apache.spark.{SparkException, Logging, SparkEnv}
+import org.apache.spark.rdd.PairRDDFunctions
 import org.apache.spark.streaming._
 
 
@@ -168,7 +169,12 @@ class JobScheduler(val ssc: StreamingContext) extends Logging {
   private class JobHandler(job: Job) extends Runnable {
     def run() {
       eventActor ! JobStarted(job)
-      job.run()
+      // Disable checks for existing output directories in jobs launched by the streaming scheduler,
+      // since we may need to write output to an existing directory during checkpoint recovery;
+      // see SPARK-4835 for more details.
+      PairRDDFunctions.disableOutputSpecValidation.withValue(true) {
+        job.run()
+      }
       eventActor ! JobCompleted(job)
     }
   }

streaming/src/test/scala/org/apache/spark/streaming/CheckpointSuite.scala

@@ -256,6 +256,45 @@ class CheckpointSuite extends TestSuiteBase {
     }
   }
 
+  test("recovery with saveAsHadoopFile inside transform operation") {
+    // Regression test for SPARK-4835.
+    //
+    // In that issue, the problem was that `saveAsHadoopFile(s)` would fail when the last batch
+    // was restarted from a checkpoint since the output directory would already exist.  However,
+    // the other saveAsHadoopFile* tests couldn't catch this because they only tested whether the
+    // output matched correctly and not whether the post-restart batch had successfully finished
+    // without throwing any errors.  The following test reproduces the same bug with a test that
+    // actually fails because the error in saveAsHadoopFile causes transform() to fail, which
+    // prevents the expected output from being written to the output stream.
+    //
+    // This is not actually a valid use of transform, but it's being used here so that we can test
+    // the fix for SPARK-4835 independently of additional test cleanup.
+    //
+    // After SPARK-5079 is addressed, should be able to remove this test since a strengthened
+    // version of the other saveAsHadoopFile* tests would prevent regressions for this issue.
+    val tempDir = Files.createTempDir()
+    try {
+      testCheckpointedOperation(
+        Seq(Seq("a", "a", "b"), Seq("", ""), Seq(), Seq("a", "a", "b"), Seq("", ""), Seq()),
+        (s: DStream[String]) => {
+          s.transform { (rdd, time) =>
+            val output = rdd.map(x => (x, 1)).reduceByKey(_ + _)
+            output.saveAsHadoopFile(
+              new File(tempDir, "result-" + time.milliseconds).getAbsolutePath,
+              classOf[Text],
+              classOf[IntWritable],
+              classOf[TextOutputFormat[Text, IntWritable]])
+            output
+          }
+        },
+        Seq(Seq(("a", 2), ("b", 1)), Seq(("", 2)), Seq(), Seq(("a", 2), ("b", 1)), Seq(("", 2)), Seq()),
+        3
+      )
+    } finally {
+      Utils.deleteRecursively(tempDir)
+    }
+  }
+
   // This tests whether the StateDStream's RDD checkpoints works correctly such
   // that the system can recover from a master failure. This assumes as reliable,
   // replayable input source - TestInputDStream.