[SPARK-3971] [MLLib] [PySpark] hotfix: Customized pickler should work…

… in cluster mode

Customized pickler should be registered before unpickling, but in executor, there is no way to register the picklers before run the tasks.

So, we need to register the picklers in the tasks itself, duplicate the javaToPython() and pythonToJava() in MLlib, call SerDe.initialize() before pickling or unpickling.

Author: Davies Liu <[email protected]>

Closes apache#2830 from davies/fix_pickle and squashes the following commits:

0c85fb9 [Davies Liu] revert the privacy change
6b94e15 [Davies Liu] use JavaConverters instead of JavaConversions
0f02050 [Davies Liu] hotfix: Customized pickler does not work in cluster

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

@@ -23,6 +23,7 @@ import java.nio.charset.Charset
 import java.util.{List => JList, ArrayList => JArrayList, Map => JMap, Collections}
 
 import scala.collection.JavaConversions._
+import scala.collection.JavaConverters._
 import scala.collection.mutable
 import scala.language.existentials
 
@@ -746,6 +747,7 @@ private[spark] object PythonRDD extends Logging {
   def pythonToJavaMap(pyRDD: JavaRDD[Array[Byte]]): JavaRDD[Map[String, _]] = {
     pyRDD.rdd.mapPartitions { iter =>
       val unpickle = new Unpickler
+      SerDeUtil.initialize()
       iter.flatMap { row =>
         unpickle.loads(row) match {
           // in case of objects are pickled in batch mode
@@ -785,7 +787,7 @@ private[spark] object PythonRDD extends Logging {
     }.toJavaRDD()
   }
 
-  private class AutoBatchedPickler(iter: Iterator[Any]) extends Iterator[Array[Byte]] {
+  private[spark] class AutoBatchedPickler(iter: Iterator[Any]) extends Iterator[Array[Byte]] {
     private val pickle = new Pickler()
     private var batch = 1
     private val buffer = new mutable.ArrayBuffer[Any]
@@ -822,11 +824,12 @@ private[spark] object PythonRDD extends Logging {
     */
   def pythonToJava(pyRDD: JavaRDD[Array[Byte]], batched: Boolean): JavaRDD[Any] = {
     pyRDD.rdd.mapPartitions { iter =>
+      SerDeUtil.initialize()
       val unpickle = new Unpickler
       iter.flatMap { row =>
         val obj = unpickle.loads(row)
         if (batched) {
-          obj.asInstanceOf[JArrayList[_]]
+          obj.asInstanceOf[JArrayList[_]].asScala
         } else {
           Seq(obj)
         }

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

@@ -29,7 +29,7 @@ import org.apache.spark.{Logging, SparkException}
 import org.apache.spark.rdd.RDD
 
 /** Utilities for serialization / deserialization between Python and Java, using Pickle. */
-private[python] object SerDeUtil extends Logging {
+private[spark] object SerDeUtil extends Logging {
   // Unpickle array.array generated by Python 2.6
   class ArrayConstructor extends net.razorvine.pickle.objects.ArrayConstructor {
     //  /* Description of types */
@@ -76,9 +76,18 @@ private[python] object SerDeUtil extends Logging {
     }
   }
 
+  private var initialized = false
+  // This should be called before trying to unpickle array.array from Python
+  // In cluster mode, this should be put in closure
   def initialize() = {
-    Unpickler.registerConstructor("array", "array", new ArrayConstructor())
+    synchronized{
+      if (!initialized) {
+        Unpickler.registerConstructor("array", "array", new ArrayConstructor())
+        initialized = true
+      }
+    }
   }
+  initialize()
 
   private def checkPickle(t: (Any, Any)): (Boolean, Boolean) = {
     val pickle = new Pickler
@@ -143,6 +152,7 @@ private[python] object SerDeUtil extends Logging {
         obj.asInstanceOf[Array[_]].length == 2
     }
     pyRDD.mapPartitions { iter =>
+      initialize()
       val unpickle = new Unpickler
       val unpickled =
         if (batchSerialized) {

mllib/src/main/scala/org/apache/spark/mllib/api/python/PythonMLLibAPI.scala

@@ -18,6 +18,7 @@
 package org.apache.spark.mllib.api.python
 
 import java.io.OutputStream
+import java.util.{ArrayList => JArrayList}
 
 import scala.collection.JavaConverters._
 import scala.language.existentials
@@ -27,6 +28,7 @@ import net.razorvine.pickle._
 
 import org.apache.spark.annotation.DeveloperApi
 import org.apache.spark.api.java.{JavaRDD, JavaSparkContext}
+import org.apache.spark.api.python.{PythonRDD, SerDeUtil}
 import org.apache.spark.mllib.classification._
 import org.apache.spark.mllib.clustering._
 import org.apache.spark.mllib.feature.Word2Vec
@@ -639,13 +641,24 @@ private[spark] object SerDe extends Serializable {
     }
   }
 
+  var initialized = false
+  // This should be called before trying to serialize any above classes
+  // In cluster mode, this should be put in the closure
   def initialize(): Unit = {
-    new DenseVectorPickler().register()
-    new DenseMatrixPickler().register()
-    new SparseVectorPickler().register()
-    new LabeledPointPickler().register()
-    new RatingPickler().register()
+    SerDeUtil.initialize()
+    synchronized {
+      if (!initialized) {
+        new DenseVectorPickler().register()
+        new DenseMatrixPickler().register()
+        new SparseVectorPickler().register()
+        new LabeledPointPickler().register()
+        new RatingPickler().register()
+        initialized = true
+      }
+    }
   }
+  // will not called in Executor automatically
+  initialize()
 
   def dumps(obj: AnyRef): Array[Byte] = {
     new Pickler().dumps(obj)
@@ -659,4 +672,33 @@ private[spark] object SerDe extends Serializable {
   def asTupleRDD(rdd: RDD[Array[Any]]): RDD[(Int, Int)] = {
     rdd.map(x => (x(0).asInstanceOf[Int], x(1).asInstanceOf[Int]))
   }
+
+  /**
+   * Convert an RDD of Java objects to an RDD of serialized Python objects, that is usable by
+   * PySpark.
+   */
+  def javaToPython(jRDD: JavaRDD[Any]): JavaRDD[Array[Byte]] = {
+    jRDD.rdd.mapPartitions { iter =>
+      initialize()  // let it called in executor
+      new PythonRDD.AutoBatchedPickler(iter)
+    }
+  }
+
+  /**
+   * Convert an RDD of serialized Python objects to RDD of objects, that is usable by PySpark.
+   */
+  def pythonToJava(pyRDD: JavaRDD[Array[Byte]], batched: Boolean): JavaRDD[Any] = {
+    pyRDD.rdd.mapPartitions { iter =>
+      initialize()  // let it called in executor
+      val unpickle = new Unpickler
+      iter.flatMap { row =>
+        val obj = unpickle.loads(row)
+        if (batched) {
+          obj.asInstanceOf[JArrayList[_]].asScala
+        } else {
+          Seq(obj)
+        }
+      }
+    }.toJavaRDD()
+  }
 }

python/pyspark/context.py

@@ -215,8 +215,6 @@ def _ensure_initialized(cls, instance=None, gateway=None):
                 SparkContext._gateway = gateway or launch_gateway()
                 SparkContext._jvm = SparkContext._gateway.jvm
                 SparkContext._writeToFile = SparkContext._jvm.PythonRDD.writeToFile
-                SparkContext._jvm.SerDeUtil.initialize()
-                SparkContext._jvm.SerDe.initialize()
 
             if instance:
                 if (SparkContext._active_spark_context and

python/pyspark/mllib/classification.py

@@ -21,7 +21,7 @@
 from numpy import array
 
 from pyspark import SparkContext, PickleSerializer
-from pyspark.mllib.linalg import SparseVector, _convert_to_vector
+from pyspark.mllib.linalg import SparseVector, _convert_to_vector, _to_java_object_rdd
 from pyspark.mllib.regression import LabeledPoint, LinearModel, _regression_train_wrapper
 
 
@@ -244,7 +244,7 @@ def train(cls, data, lambda_=1.0):
         :param lambda_: The smoothing parameter
         """
         sc = data.context
-        jlist = sc._jvm.PythonMLLibAPI().trainNaiveBayes(data._to_java_object_rdd(), lambda_)
+        jlist = sc._jvm.PythonMLLibAPI().trainNaiveBayes(_to_java_object_rdd(data), lambda_)
         labels, pi, theta = PickleSerializer().loads(str(sc._jvm.SerDe.dumps(jlist)))
         return NaiveBayesModel(labels.toArray(), pi.toArray(), numpy.array(theta))
 

python/pyspark/mllib/clustering.py

@@ -17,7 +17,7 @@
 
 from pyspark import SparkContext
 from pyspark.serializers import PickleSerializer, AutoBatchedSerializer
-from pyspark.mllib.linalg import SparseVector, _convert_to_vector
+from pyspark.mllib.linalg import SparseVector, _convert_to_vector, _to_java_object_rdd
 
 __all__ = ['KMeansModel', 'KMeans']
 
@@ -85,7 +85,7 @@ def train(cls, rdd, k, maxIterations=100, runs=1, initializationMode="k-means||"
         # cache serialized data to avoid objects over head in JVM
         cached = rdd.map(_convert_to_vector)._reserialize(AutoBatchedSerializer(ser)).cache()
         model = sc._jvm.PythonMLLibAPI().trainKMeansModel(
-            cached._to_java_object_rdd(), k, maxIterations, runs, initializationMode)
+            _to_java_object_rdd(cached), k, maxIterations, runs, initializationMode)
         bytes = sc._jvm.SerDe.dumps(model.clusterCenters())
         centers = ser.loads(str(bytes))
         return KMeansModel([c.toArray() for c in centers])

python/pyspark/mllib/feature.py

@@ -19,8 +19,7 @@
 Python package for feature in MLlib.
 """
 from pyspark.serializers import PickleSerializer, AutoBatchedSerializer
-
-from pyspark.mllib.linalg import _convert_to_vector
+from pyspark.mllib.linalg import _convert_to_vector, _to_java_object_rdd
 
 __all__ = ['Word2Vec', 'Word2VecModel']
 
@@ -176,7 +175,7 @@ def fit(self, data):
         seed = self.seed
 
         model = sc._jvm.PythonMLLibAPI().trainWord2Vec(
-            data._to_java_object_rdd(), vectorSize,
+            _to_java_object_rdd(data), vectorSize,
             learningRate, numPartitions, numIterations, seed)
         return Word2VecModel(sc, model)
 

python/pyspark/mllib/linalg.py

@@ -29,6 +29,8 @@
 
 import numpy as np
 
+from pyspark.serializers import AutoBatchedSerializer, PickleSerializer
+
 __all__ = ['Vector', 'DenseVector', 'SparseVector', 'Vectors']
 
 
@@ -50,6 +52,17 @@ def fast_pickle_array(ar):
     _have_scipy = False
 
 
+# this will call the MLlib version of pythonToJava()
+def _to_java_object_rdd(rdd):
+    """ Return an JavaRDD of Object by unpickling
+
+    It will convert each Python object into Java object by Pyrolite, whenever the
+    RDD is serialized in batch or not.
+    """
+    rdd = rdd._reserialize(AutoBatchedSerializer(PickleSerializer()))
+    return rdd.ctx._jvm.SerDe.pythonToJava(rdd._jrdd, True)
+
+
 def _convert_to_vector(l):
     if isinstance(l, Vector):
         return l

python/pyspark/mllib/random.py

@@ -32,7 +32,7 @@ def serialize(f):
     @wraps(f)
     def func(sc, *a, **kw):
         jrdd = f(sc, *a, **kw)
-        return RDD(sc._jvm.PythonRDD.javaToPython(jrdd), sc,
+        return RDD(sc._jvm.SerDe.javaToPython(jrdd), sc,
                    BatchedSerializer(PickleSerializer(), 1024))
     return func
 

python/pyspark/mllib/recommendation.py

@@ -18,6 +18,7 @@
 from pyspark import SparkContext
 from pyspark.serializers import PickleSerializer, AutoBatchedSerializer
 from pyspark.rdd import RDD
+from pyspark.mllib.linalg import _to_java_object_rdd
 
 __all__ = ['MatrixFactorizationModel', 'ALS']
 
@@ -77,9 +78,9 @@ def predictAll(self, user_product):
             first = tuple(map(int, first))
         assert all(type(x) is int for x in first), "user and product in user_product shoul be int"
         sc = self._context
-        tuplerdd = sc._jvm.SerDe.asTupleRDD(user_product._to_java_object_rdd().rdd())
+        tuplerdd = sc._jvm.SerDe.asTupleRDD(_to_java_object_rdd(user_product).rdd())
         jresult = self._java_model.predict(tuplerdd).toJavaRDD()
-        return RDD(sc._jvm.PythonRDD.javaToPython(jresult), sc,
+        return RDD(sc._jvm.SerDe.javaToPython(jresult), sc,
                    AutoBatchedSerializer(PickleSerializer()))
 
 
@@ -97,7 +98,7 @@ def _prepare(cls, ratings):
         # serialize them by AutoBatchedSerializer before cache to reduce the
         # objects overhead in JVM
         cached = ratings._reserialize(AutoBatchedSerializer(PickleSerializer())).cache()
-        return cached._to_java_object_rdd()
+        return _to_java_object_rdd(cached)
 
     @classmethod
     def train(cls, ratings, rank, iterations=5, lambda_=0.01, blocks=-1):

python/pyspark/mllib/regression.py

@@ -19,8 +19,8 @@
 from numpy import array
 
 from pyspark import SparkContext
-from pyspark.mllib.linalg import SparseVector, _convert_to_vector
 from pyspark.serializers import PickleSerializer, AutoBatchedSerializer
+from pyspark.mllib.linalg import SparseVector, _convert_to_vector, _to_java_object_rdd
 
 __all__ = ['LabeledPoint', 'LinearModel', 'LinearRegressionModel', 'RidgeRegressionModel',
            'LinearRegressionWithSGD', 'LassoWithSGD', 'RidgeRegressionWithSGD']
@@ -131,7 +131,7 @@ def _regression_train_wrapper(sc, train_func, modelClass, data, initial_weights)
     # use AutoBatchedSerializer before cache to reduce the memory
     # overhead in JVM
     cached = data._reserialize(AutoBatchedSerializer(ser)).cache()
-    ans = train_func(cached._to_java_object_rdd(), initial_bytes)
+    ans = train_func(_to_java_object_rdd(cached), initial_bytes)
     assert len(ans) == 2, "JVM call result had unexpected length"
     weights = ser.loads(str(ans[0]))
     return modelClass(weights, ans[1])

python/pyspark/mllib/stat.py

@@ -22,6 +22,7 @@
 from functools import wraps
 
 from pyspark import PickleSerializer
+from pyspark.mllib.linalg import _to_java_object_rdd
 
 
 __all__ = ['MultivariateStatisticalSummary', 'Statistics']
@@ -106,7 +107,7 @@ def colStats(rdd):
         array([ 2.,  0.,  0., -2.])
         """
         sc = rdd.ctx
-        jrdd = rdd._to_java_object_rdd()
+        jrdd = _to_java_object_rdd(rdd)
         cStats = sc._jvm.PythonMLLibAPI().colStats(jrdd)
         return MultivariateStatisticalSummary(sc, cStats)
 
@@ -162,14 +163,14 @@ def corr(x, y=None, method=None):
         if type(y) == str:
             raise TypeError("Use 'method=' to specify method name.")
 
-        jx = x._to_java_object_rdd()
+        jx = _to_java_object_rdd(x)
         if not y:
             resultMat = sc._jvm.PythonMLLibAPI().corr(jx, method)
             bytes = sc._jvm.SerDe.dumps(resultMat)
             ser = PickleSerializer()
             return ser.loads(str(bytes)).toArray()
         else:
-            jy = y._to_java_object_rdd()
+            jy = _to_java_object_rdd(y)
             return sc._jvm.PythonMLLibAPI().corr(jx, jy, method)
 
 

python/pyspark/mllib/tree.py

@@ -19,7 +19,7 @@
 
 from pyspark import SparkContext, RDD
 from pyspark.serializers import BatchedSerializer, PickleSerializer
-from pyspark.mllib.linalg import Vector, _convert_to_vector
+from pyspark.mllib.linalg import Vector, _convert_to_vector, _to_java_object_rdd
 from pyspark.mllib.regression import LabeledPoint
 
 __all__ = ['DecisionTreeModel', 'DecisionTree']
@@ -61,8 +61,8 @@ def predict(self, x):
                 return self._sc.parallelize([])
             if not isinstance(first[0], Vector):
                 x = x.map(_convert_to_vector)
-            jPred = self._java_model.predict(x._to_java_object_rdd()).toJavaRDD()
-            jpyrdd = self._sc._jvm.PythonRDD.javaToPython(jPred)
+            jPred = self._java_model.predict(_to_java_object_rdd(x)).toJavaRDD()
+            jpyrdd = self._sc._jvm.SerDe.javaToPython(jPred)
             return RDD(jpyrdd, self._sc, BatchedSerializer(ser, 1024))
 
         else:
@@ -104,7 +104,7 @@ def _train(data, type, numClasses, categoricalFeaturesInfo,
         first = data.first()
         assert isinstance(first, LabeledPoint), "the data should be RDD of LabeledPoint"
         sc = data.context
-        jrdd = data._to_java_object_rdd()
+        jrdd = _to_java_object_rdd(data)
         cfiMap = MapConverter().convert(categoricalFeaturesInfo,
                                         sc._gateway._gateway_client)
         model = sc._jvm.PythonMLLibAPI().trainDecisionTreeModel(

python/pyspark/mllib/util.py

@@ -19,7 +19,7 @@
 import warnings
 
 from pyspark.rdd import RDD
-from pyspark.serializers import BatchedSerializer, PickleSerializer
+from pyspark.serializers import AutoBatchedSerializer, PickleSerializer
 from pyspark.mllib.linalg import Vectors, SparseVector, _convert_to_vector
 from pyspark.mllib.regression import LabeledPoint
 
@@ -174,8 +174,8 @@ def loadLabeledPoints(sc, path, minPartitions=None):
         """
         minPartitions = minPartitions or min(sc.defaultParallelism, 2)
         jrdd = sc._jvm.PythonMLLibAPI().loadLabeledPoints(sc._jsc, path, minPartitions)
-        jpyrdd = sc._jvm.PythonRDD.javaToPython(jrdd)
-        return RDD(jpyrdd, sc, BatchedSerializer(PickleSerializer()))
+        jpyrdd = sc._jvm.SerDe.javaToPython(jrdd)
+        return RDD(jpyrdd, sc, AutoBatchedSerializer(PickleSerializer()))
 
 
 def _test():