[SPARK-4835] Disable OutputSpec validation for streaming jobs.

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}
@@ -961,7 +962,8 @@ class PairRDDFunctions[K, V](self: RDD[(K, V)])
     val outfmt = job.getOutputFormatClass
     val jobFormat = outfmt.newInstance
 
-    if (self.conf.getBoolean("spark.hadoop.validateOutputSpecs", true)) {
+    val validationDisabled = PairRDDFunctions.disableOutputSpecValidation.value
+    if (!validationDisabled && self.conf.getBoolean("spark.hadoop.validateOutputSpecs", true)) {
       // FileOutputFormat ignores the filesystem parameter
       jobFormat.checkOutputSpecs(job)
     }
@@ -1039,7 +1041,8 @@ 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)) {
+    val validationDisabled = PairRDDFunctions.disableOutputSpecValidation.value
+    if (!validationDisabled && self.conf.getBoolean("spark.hadoop.validateOutputSpecs", true)) {
       // FileOutputFormat ignores the filesystem parameter
       val ignoredFs = FileSystem.get(hadoopConf)
       hadoopConf.getOutputFormat.checkOutputSpecs(ignoredFs, hadoopConf)
@@ -1118,4 +1121,10 @@ class PairRDDFunctions[K, V](self: RDD[(K, V)])
 
 private[spark] object PairRDDFunctions {
   val RECORDS_BETWEEN_BYTES_WRITTEN_METRIC_UPDATES = 256
+  /**
+   * Used by Spark Streaming in order to bypass the `spark.hadoop.validateOutputSpecs` checks
+   * for save actions launched by Spark Streaming, since the validation may break Spark Streaming's
+   * ability to recover from checkpoints.  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 launched by the Spark Streaming scheduler, since data may need
+    to be written to a pre-existing output directory during checkpoint recovery.</td>
 </tr>
 <tr>
     <td><code>spark.hadoop.cloneConf</code></td>

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)
     }
   }