[SPARK-6368][SQL] Build a specialized serializer for Exchange operator.

JIRA: https://issues.apache.org/jira/browse/SPARK-6368

Author: Yin Huai <[email protected]>

Closes apache#5497 from yhuai/serializer2 and squashes the following commits:

da562c5 [Yin Huai] Merge remote-tracking branch 'upstream/master' into serializer2
50e0c3d [Yin Huai] When no filed is emitted to shuffle, use SparkSqlSerializer for now.
9f1ed92 [Yin Huai] Merge remote-tracking branch 'upstream/master' into serializer2
6d07678 [Yin Huai] Address comments.
4273b8c [Yin Huai] Enabled SparkSqlSerializer2.
09e587a [Yin Huai] Remove TODO.
791b96a [Yin Huai] Use UTF8String.
60a1487 [Yin Huai] Merge remote-tracking branch 'upstream/master' into serializer2
3e09655 [Yin Huai] Use getAs for Date column.
43b9fb4 [Yin Huai] Test.
8297732 [Yin Huai] Fix test.
c9373c8 [Yin Huai] Support DecimalType.
2379eeb [Yin Huai] ASF header.
39704ab [Yin Huai] Specialized serializer for Exchange.

Author: yhuai

sql/core/src/main/scala/org/apache/spark/sql/SQLConf.scala

@@ -64,6 +64,8 @@ private[spark] object SQLConf {
   // Set to false when debugging requires the ability to look at invalid query plans.
   val DATAFRAME_EAGER_ANALYSIS = "spark.sql.eagerAnalysis"
 
+  val USE_SQL_SERIALIZER2 = "spark.sql.useSerializer2"
+
   object Deprecated {
     val MAPRED_REDUCE_TASKS = "mapred.reduce.tasks"
   }
@@ -147,6 +149,8 @@ private[sql] class SQLConf extends Serializable {
    */
   private[spark] def codegenEnabled: Boolean = getConf(CODEGEN_ENABLED, "false").toBoolean
 
+  private[spark] def useSqlSerializer2: Boolean = getConf(USE_SQL_SERIALIZER2, "true").toBoolean
+
   /**
    * Upper bound on the sizes (in bytes) of the tables qualified for the auto conversion to
    * a broadcast value during the physical executions of join operations.  Setting this to -1

sql/core/src/main/scala/org/apache/spark/sql/execution/Exchange.scala

@@ -19,13 +19,15 @@ package org.apache.spark.sql.execution
 
 import org.apache.spark.annotation.DeveloperApi
 import org.apache.spark.shuffle.sort.SortShuffleManager
-import org.apache.spark.{SparkEnv, HashPartitioner, RangePartitioner, SparkConf}
+import org.apache.spark.{SparkEnv, HashPartitioner, RangePartitioner}
 import org.apache.spark.rdd.{RDD, ShuffledRDD}
+import org.apache.spark.serializer.Serializer
 import org.apache.spark.sql.{SQLContext, Row}
 import org.apache.spark.sql.catalyst.errors.attachTree
 import org.apache.spark.sql.catalyst.expressions._
 import org.apache.spark.sql.catalyst.plans.physical._
 import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.types.DataType
 import org.apache.spark.util.MutablePair
 
 object Exchange {
@@ -77,9 +79,48 @@ case class Exchange(
     }
   }
 
-  override def execute(): RDD[Row] = attachTree(this , "execute") {
-    lazy val sparkConf = child.sqlContext.sparkContext.getConf
+  @transient private lazy val sparkConf = child.sqlContext.sparkContext.getConf
+
+  def serializer(
+      keySchema: Array[DataType],
+      valueSchema: Array[DataType],
+      numPartitions: Int): Serializer = {
+    // In ExternalSorter's spillToMergeableFile function, key-value pairs are written out
+    // through write(key) and then write(value) instead of write((key, value)). Because
+    // SparkSqlSerializer2 assumes that objects passed in are Product2, we cannot safely use
+    // it when spillToMergeableFile in ExternalSorter will be used.
+    // So, we will not use SparkSqlSerializer2 when
+    //  - Sort-based shuffle is enabled and the number of reducers (numPartitions) is greater
+    //     then the bypassMergeThreshold; or
+    //  - newOrdering is defined.
+    val cannotUseSqlSerializer2 =
+      (sortBasedShuffleOn && numPartitions > bypassMergeThreshold) || newOrdering.nonEmpty
+
+    // It is true when there is no field that needs to be write out.
+    // For now, we will not use SparkSqlSerializer2 when noField is true.
+    val noField =
+      (keySchema == null || keySchema.length == 0) &&
+      (valueSchema == null || valueSchema.length == 0)
+
+    val useSqlSerializer2 =
+        child.sqlContext.conf.useSqlSerializer2 &&   // SparkSqlSerializer2 is enabled.
+        !cannotUseSqlSerializer2 &&                  // Safe to use Serializer2.
+        SparkSqlSerializer2.support(keySchema) &&    // The schema of key is supported.
+        SparkSqlSerializer2.support(valueSchema) &&  // The schema of value is supported.
+        !noField
+
+    val serializer = if (useSqlSerializer2) {
+      logInfo("Using SparkSqlSerializer2.")
+      new SparkSqlSerializer2(keySchema, valueSchema)
+    } else {
+      logInfo("Using SparkSqlSerializer.")
+      new SparkSqlSerializer(sparkConf)
+    }
+
+    serializer
+  }
 
+  override def execute(): RDD[Row] = attachTree(this , "execute") {
     newPartitioning match {
       case HashPartitioning(expressions, numPartitions) =>
         // TODO: Eliminate redundant expressions in grouping key and value.
@@ -111,7 +152,10 @@ case class Exchange(
           } else {
             new ShuffledRDD[Row, Row, Row](rdd, part)
           }
-        shuffled.setSerializer(new SparkSqlSerializer(sparkConf))
+        val keySchema = expressions.map(_.dataType).toArray
+        val valueSchema = child.output.map(_.dataType).toArray
+        shuffled.setSerializer(serializer(keySchema, valueSchema, numPartitions))
+
         shuffled.map(_._2)
 
       case RangePartitioning(sortingExpressions, numPartitions) =>
@@ -134,7 +178,9 @@ case class Exchange(
           } else {
             new ShuffledRDD[Row, Null, Null](rdd, part)
           }
-        shuffled.setSerializer(new SparkSqlSerializer(sparkConf))
+        val keySchema = child.output.map(_.dataType).toArray
+        shuffled.setSerializer(serializer(keySchema, null, numPartitions))
+
         shuffled.map(_._1)
 
       case SinglePartition =>
@@ -152,7 +198,8 @@ case class Exchange(
         }
         val partitioner = new HashPartitioner(1)
         val shuffled = new ShuffledRDD[Null, Row, Row](rdd, partitioner)
-        shuffled.setSerializer(new SparkSqlSerializer(sparkConf))
+        val valueSchema = child.output.map(_.dataType).toArray
+        shuffled.setSerializer(serializer(null, valueSchema, 1))
         shuffled.map(_._2)
 
       case _ => sys.error(s"Exchange not implemented for $newPartitioning")