[SPARK-14267] [SQL] [PYSPARK] execute multiple Python UDFs within single batch

core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala

@@ -59,7 +59,7 @@ private[spark] class PythonRDD(
   val asJavaRDD: JavaRDD[Array[Byte]] = JavaRDD.fromRDD(this)
 
   override def compute(split: Partition, context: TaskContext): Iterator[Array[Byte]] = {
-    val runner = new PythonRunner(Seq(func), bufferSize, reuse_worker, false)
+    val runner = PythonRunner(func, bufferSize, reuse_worker)
     runner.compute(firstParent.iterator(split, context), split.index, context)
   }
 }
@@ -77,22 +77,30 @@ private[spark] case class PythonFunction(
     broadcastVars: JList[Broadcast[PythonBroadcast]],
     accumulator: Accumulator[JList[Array[Byte]]])
 
+
+object PythonRunner {
+  def apply(func: PythonFunction, bufferSize: Int, reuse_worker: Boolean): PythonRunner = {
+    new PythonRunner(Seq(Seq(func)), bufferSize, reuse_worker, false, Seq(1))
+  }
+}
+
 /**
- * A helper class to run Python UDFs in Spark.
+ * A helper class to run Python mapPartition/UDFs in Spark.
  */
 private[spark] class PythonRunner(
-    funcs: Seq[PythonFunction],
+    funcs: Seq[Seq[PythonFunction]],
     bufferSize: Int,
     reuse_worker: Boolean,
-    rowBased: Boolean)
+    isUDF: Boolean,
+    numArgs: Seq[Int])
   extends Logging {
 
   // All the Python functions should have the same exec, version and envvars.
-  private val envVars = funcs.head.envVars
-  private val pythonExec = funcs.head.pythonExec
-  private val pythonVer = funcs.head.pythonVer
+  private val envVars = funcs.head.head.envVars
+  private val pythonExec = funcs.head.head.pythonExec
+  private val pythonVer = funcs.head.head.pythonVer
 
-  private val accumulator = funcs.head.accumulator // TODO: support accumulator in multiple UDF
+  private val accumulator = funcs.head.head.accumulator // TODO: support accumulator in multiple UDF
 
   def compute(
       inputIterator: Iterator[_],
@@ -232,8 +240,8 @@ private[spark] class PythonRunner(
 
     @volatile private var _exception: Exception = null
 
-    private val pythonIncludes = funcs.flatMap(_.pythonIncludes.asScala).toSet
-    private val broadcastVars = funcs.flatMap(_.broadcastVars.asScala)
+    private val pythonIncludes = funcs.flatMap(_.flatMap(_.pythonIncludes.asScala)).toSet
+    private val broadcastVars = funcs.flatMap(_.flatMap(_.broadcastVars.asScala))
 
     setDaemon(true)
 
@@ -284,11 +292,22 @@ private[spark] class PythonRunner(
         }
         dataOut.flush()
         // Serialized command:
-        dataOut.writeInt(if (rowBased) 1 else 0)
-        dataOut.writeInt(funcs.length)
-        funcs.foreach { f =>
-          dataOut.writeInt(f.command.length)
-          dataOut.write(f.command)
+        if (isUDF) {
+          dataOut.writeInt(1)
+          dataOut.writeInt(funcs.length)
+          funcs.zip(numArgs).foreach { case (fs, numArg) =>
+            dataOut.writeInt(numArg)
+            dataOut.writeInt(fs.length)
+            fs.foreach { f =>
+              dataOut.writeInt(f.command.length)
+              dataOut.write(f.command)
+            }
+          }
+        } else {
+          dataOut.writeInt(0)
+          val command = funcs.head.head.command
+          dataOut.writeInt(command.length)
+          dataOut.write(command)
         }
         // Data values
         PythonRDD.writeIteratorToStream(inputIterator, dataOut)

python/pyspark/sql/functions.py

@@ -1649,8 +1649,7 @@ def sort_array(col, asc=True):
 # ---------------------------- User Defined Function ----------------------------------
 
 def _wrap_function(sc, func, returnType):
-    ser = AutoBatchedSerializer(PickleSerializer())
-    command = (func, returnType, ser)
+    command = (func, returnType)
     pickled_command, broadcast_vars, env, includes = _prepare_for_python_RDD(sc, command)
     return sc._jvm.PythonFunction(bytearray(pickled_command), env, includes, sc.pythonExec,
                                   sc.pythonVer, broadcast_vars, sc._javaAccumulator)

python/pyspark/sql/tests.py

@@ -305,7 +305,7 @@ def test_udf2(self):
         [res] = self.sqlCtx.sql("SELECT strlen(a) FROM test WHERE strlen(a) > 1").collect()
         self.assertEqual(4, res[0])
 
-    def test_chained_python_udf(self):
+    def test_chained_udf(self):
         self.sqlCtx.registerFunction("double", lambda x: x + x, IntegerType())
         [row] = self.sqlCtx.sql("SELECT double(1)").collect()
         self.assertEqual(row[0], 2)
@@ -314,6 +314,16 @@ def test_chained_python_udf(self):
         [row] = self.sqlCtx.sql("SELECT double(double(1) + 1)").collect()
         self.assertEqual(row[0], 6)
 
+    def test_multiple_udfs(self):
+        self.sqlCtx.registerFunction("double", lambda x: x * 2, IntegerType())
+        [row] = self.sqlCtx.sql("SELECT double(1), double(2)").collect()
+        self.assertEqual(tuple(row), (2, 4))
+        [row] = self.sqlCtx.sql("SELECT double(double(1)), double(double(2) + 2)").collect()
+        self.assertEqual(tuple(row), (4, 12))
+        self.sqlCtx.registerFunction("add", lambda x, y: x + y, IntegerType())
+        [row] = self.sqlCtx.sql("SELECT double(add(1, 2)), add(double(2), 1)").collect()
+        self.assertEqual(tuple(row), (6, 5))
+
     def test_udf_with_array_type(self):
         d = [Row(l=list(range(3)), d={"key": list(range(5))})]
         rdd = self.sc.parallelize(d)

python/pyspark/worker.py

@@ -29,7 +29,7 @@
 from pyspark.broadcast import Broadcast, _broadcastRegistry
 from pyspark.files import SparkFiles
 from pyspark.serializers import write_with_length, write_int, read_long, \
-    write_long, read_int, SpecialLengths, UTF8Deserializer, PickleSerializer
+    write_long, read_int, SpecialLengths, UTF8Deserializer, PickleSerializer, AutoBatchedSerializer
 from pyspark import shuffle
 
 pickleSer = PickleSerializer()
@@ -59,7 +59,49 @@ def read_command(serializer, file):
 
 def chain(f, g):
     """chain two function together """
-    return lambda x: g(f(x))
+    return lambda *a: g(f(*a))
+
+
+def wrap_udf(f, return_type):
+    return lambda *a: return_type.toInternal(f(*a))
+
+
+def read_single_udf(pickleSer, infile):
+    num_arg = read_int(infile)
+    row_func = None
+    for i in range(read_int(infile)):
+        f, return_type = read_command(pickleSer, infile)
+        if row_func is None:
+            row_func = f
+        else:
+            row_func = chain(row_func, f)
+    # the last returnType will be the return type of UDF
+    return num_arg, wrap_udf(row_func, return_type)
+
+
+def read_udfs(pickleSer, infile):
+    num_udfs = read_int(infile)
+    udfs = []
+    offset = 0
+    for i in range(num_udfs):
+        num_arg, udf = read_single_udf(pickleSer, infile)
+        udfs.append((offset, offset + num_arg, udf))
+        offset += num_arg
+
+    if num_udfs == 1:
+        udf = udfs[0][2]
+
+        # fast path for single UDF
+        def mapper(args):
+            return udf(*args)
+    else:
+        def mapper(args):
+            return tuple(udf(*args[start:end]) for start, end, udf in udfs)
+
+    func = lambda _, it: map(mapper, it)
+    ser = AutoBatchedSerializer(PickleSerializer())
+    # profiling is not supported for UDF
+    return func, None, ser, ser
 
 
 def main(infile, outfile):
@@ -107,21 +149,10 @@ def main(infile, outfile):
                 _broadcastRegistry.pop(bid)
 
         _accumulatorRegistry.clear()
-        row_based = read_int(infile)
-        num_commands = read_int(infile)
-        if row_based:
-            profiler = None  # profiling is not supported for UDF
-            row_func = None
-            for i in range(num_commands):
-                f, returnType, deserializer = read_command(pickleSer, infile)
-                if row_func is None:
-                    row_func = f
-                else:
-                    row_func = chain(row_func, f)
-            serializer = deserializer
-            func = lambda _, it: map(lambda x: returnType.toInternal(row_func(*x)), it)
+        is_udf = read_int(infile)
+        if is_udf:
+            func, profiler, deserializer, serializer = read_udfs(pickleSer, infile)
         else:
-            assert num_commands == 1
             func, profiler, deserializer, serializer = read_command(pickleSer, infile)
 
         init_time = time.time()

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

@@ -426,8 +426,8 @@ private[sql] abstract class SparkStrategies extends QueryPlanner[SparkPlan] {
       case logical.RepartitionByExpression(expressions, child, nPartitions) =>
         exchange.ShuffleExchange(HashPartitioning(
           expressions, nPartitions.getOrElse(numPartitions)), planLater(child)) :: Nil
-      case e @ python.EvaluatePython(udf, child, _) =>
-        python.BatchPythonEvaluation(udf, e.output, planLater(child)) :: Nil
+      case e @ python.EvaluatePython(udfs, child, _) =>
+        python.BatchPythonEvaluation(udfs, e.output, planLater(child)) :: Nil
       case LogicalRDD(output, rdd) => PhysicalRDD(output, rdd, "ExistingRDD") :: Nil
       case BroadcastHint(child) => planLater(child) :: Nil
       case _ => Nil

sql/core/src/main/scala/org/apache/spark/sql/execution/python/BatchPythonEvaluation.scala

@@ -40,7 +40,7 @@ import org.apache.spark.sql.types.{StructField, StructType}
  * we drain the queue to find the original input row. Note that if the Python process is way too
  * slow, this could lead to the queue growing unbounded and eventually run out of memory.
  */
-case class BatchPythonEvaluation(udf: PythonUDF, output: Seq[Attribute], child: SparkPlan)
+case class BatchPythonEvaluation(udfs: Seq[PythonUDF], output: Seq[Attribute], child: SparkPlan)
   extends SparkPlan {
 
   def children: Seq[SparkPlan] = child :: Nil
@@ -69,11 +69,14 @@ case class BatchPythonEvaluation(udf: PythonUDF, output: Seq[Attribute], child:
       // combine input with output from Python.
       val queue = new java.util.concurrent.ConcurrentLinkedQueue[InternalRow]()
 
-      val (pyFuncs, children) = collectFunctions(udf)
+      val (pyFuncs, children) = udfs.map(collectFunctions).unzip
+      val numArgs = children.map(_.length)
 
       val pickle = new Pickler
-      val currentRow = newMutableProjection(children, child.output)()
-      val fields = children.map(_.dataType)
+      // flatten all the arguments
+      val allChildren = children.flatMap(x => x)
+      val currentRow = newMutableProjection(allChildren, child.output)()
+      val fields = allChildren.map(_.dataType)
       val schema = new StructType(fields.map(t => new StructField("", t, true)).toArray)
 
       // Input iterator to Python: input rows are grouped so we send them in batches to Python.
@@ -89,19 +92,30 @@ case class BatchPythonEvaluation(udf: PythonUDF, output: Seq[Attribute], child:
       val context = TaskContext.get()
 
       // Output iterator for results from Python.
-      val outputIterator = new PythonRunner(pyFuncs, bufferSize, reuseWorker, true)
+      val outputIterator = new PythonRunner(pyFuncs, bufferSize, reuseWorker, true, numArgs)
         .compute(inputIterator, context.partitionId(), context)
 
       val unpickle = new Unpickler
-      val row = new GenericMutableRow(1)
+      val mutableRow = new GenericMutableRow(1)
       val joined = new JoinedRow
+      val resultType = if (udfs.length == 1) {
+        udfs.head.dataType
+      } else {
+        StructType(udfs.map(u => StructField("", u.dataType, u.nullable)))
+      }
       val resultProj = UnsafeProjection.create(output, output)
 
       outputIterator.flatMap { pickedResult =>
         val unpickledBatch = unpickle.loads(pickedResult)
         unpickledBatch.asInstanceOf[java.util.ArrayList[Any]].asScala
       }.map { result =>
-        row(0) = EvaluatePython.fromJava(result, udf.dataType)
+        val row = if (udfs.length == 1) {
+          // fast path for single UDF
+          mutableRow(0) = EvaluatePython.fromJava(result, resultType)
+          mutableRow
+        } else {
+          EvaluatePython.fromJava(result, resultType).asInstanceOf[InternalRow]
+        }
         resultProj(joined(queue.poll(), row))
       }
     }

sql/core/src/main/scala/org/apache/spark/sql/execution/python/EvaluatePython.scala

@@ -36,24 +36,28 @@ import org.apache.spark.sql.types._
 import org.apache.spark.unsafe.types.UTF8String
 
 /**
- * Evaluates a [[PythonUDF]], appending the result to the end of the input tuple.
+ * Evaluates a list of [[PythonUDF]], appending the result to the end of the input tuple.
  */
 case class EvaluatePython(
-    udf: PythonUDF,
+    udfs: Seq[PythonUDF],
     child: LogicalPlan,
-    resultAttribute: AttributeReference)
+    resultAttribute: Seq[AttributeReference])
   extends logical.UnaryNode {
 
-  def output: Seq[Attribute] = child.output :+ resultAttribute
+  def output: Seq[Attribute] = child.output ++ resultAttribute
 
   // References should not include the produced attribute.
-  override def references: AttributeSet = udf.references
+  override def references: AttributeSet = AttributeSet(udfs.flatMap(_.references))
 }
 
 
 object EvaluatePython {
-  def apply(udf: PythonUDF, child: LogicalPlan): EvaluatePython =
-    new EvaluatePython(udf, child, AttributeReference("pythonUDF", udf.dataType)())
+  def apply(udfs: Seq[PythonUDF], child: LogicalPlan): EvaluatePython = {
+    val resultAttrs = udfs.zipWithIndex.map { case (u, i) =>
+      AttributeReference(s"pythonUDF$i", u.dataType)()
+    }
+    new EvaluatePython(udfs, child, resultAttrs)
+  }
 
   def takeAndServe(df: DataFrame, n: Int): Int = {
     registerPicklers()