[SPARK-4531] [MLlib] cache serialized java object

The Pyrolite is pretty slow (comparing to the adhoc serializer in 1.1), it cause much performance regression in 1.2, because we cache the serialized Python object in JVM, deserialize them into Java object in each step.

This PR change to cache the deserialized JavaRDD instead of PythonRDD to avoid the deserialization of Pyrolite. It should have similar memory usage as before, but much faster.

Author: Davies Liu <davies@databricks.com>

Closes #3397 from davies/cache and squashes the following commits:

7f6e6ce [Davies Liu] Update -> Updater
4b52edd [Davies Liu] using named argument
63b984e [Davies Liu] fix
7da0332 [Davies Liu] add unpersist()
dff33e1 [Davies Liu] address comments
c2bdfc2 [Davies Liu] refactor
d572f00 [Davies Liu] Merge branch 'master' into cache
f1063e1 [Davies Liu] cache serialized java object

(cherry picked from commit ce95bd8e130b2c7688b94be40683bdd90d86012d)
Signed-off-by: Xiangrui Meng <meng@databricks.com>

4 files changed, 8 insertions, 13 deletions

diff --git a/python/pyspark/mllib/clustering.py b/python/pyspark/mllib/clustering.py
index fe4c4cc509..e2492eef5b 100644
--- a/python/pyspark/mllib/clustering.py
+++ b/python/pyspark/mllib/clustering.py
@@ -16,7 +16,7 @@
 #
 
 from pyspark import SparkContext
-from pyspark.mllib.common import callMLlibFunc, callJavaFunc, _to_java_object_rdd
+from pyspark.mllib.common import callMLlibFunc, callJavaFunc
 from pyspark.mllib.linalg import SparseVector, _convert_to_vector
 
 __all__ = ['KMeansModel', 'KMeans']
@@ -80,10 +80,8 @@ class KMeans(object):
     @classmethod
     def train(cls, rdd, k, maxIterations=100, runs=1, initializationMode="k-means||"):
         """Train a k-means clustering model."""
-        # cache serialized data to avoid objects over head in JVM
-        jcached = _to_java_object_rdd(rdd.map(_convert_to_vector), cache=True)
-        model = callMLlibFunc("trainKMeansModel", jcached, k, maxIterations, runs,
-                              initializationMode)
+        model = callMLlibFunc("trainKMeansModel", rdd.map(_convert_to_vector), k, maxIterations,
+                              runs, initializationMode)
         centers = callJavaFunc(rdd.context, model.clusterCenters)
         return KMeansModel([c.toArray() for c in centers])
 
diff --git a/python/pyspark/mllib/common.py b/python/pyspark/mllib/common.py
index c6149fe391..33c49e2399 100644
--- a/python/pyspark/mllib/common.py
+++ b/python/pyspark/mllib/common.py
@@ -54,15 +54,13 @@ _picklable_classes = [
 
 
 # this will call the MLlib version of pythonToJava()
-def _to_java_object_rdd(rdd, cache=False):
+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()))
-    if cache:
-        rdd.cache()
     return rdd.ctx._jvm.SerDe.pythonToJava(rdd._jrdd, True)
 
 
diff --git a/python/pyspark/mllib/recommendation.py b/python/pyspark/mllib/recommendation.py
index 2bcbf2aaf8..97ec74eda0 100644
--- a/python/pyspark/mllib/recommendation.py
+++ b/python/pyspark/mllib/recommendation.py
@@ -19,7 +19,7 @@ from collections import namedtuple
 
 from pyspark import SparkContext
 from pyspark.rdd import RDD
-from pyspark.mllib.common import JavaModelWrapper, callMLlibFunc, _to_java_object_rdd
+from pyspark.mllib.common import JavaModelWrapper, callMLlibFunc
 
 __all__ = ['MatrixFactorizationModel', 'ALS', 'Rating']
 
@@ -110,7 +110,7 @@ class ALS(object):
                 ratings = ratings.map(lambda x: Rating(*x))
             else:
                 raise ValueError("rating should be RDD of Rating or tuple/list")
-        return _to_java_object_rdd(ratings, True)
+        return ratings
 
     @classmethod
     def train(cls, ratings, rank, iterations=5, lambda_=0.01, blocks=-1, nonnegative=False,
diff --git a/python/pyspark/mllib/regression.py b/python/pyspark/mllib/regression.py
index f4f5e615fa..210060140f 100644
--- a/python/pyspark/mllib/regression.py
+++ b/python/pyspark/mllib/regression.py
@@ -18,7 +18,7 @@
 import numpy as np
 from numpy import array
 
-from pyspark.mllib.common import callMLlibFunc, _to_java_object_rdd
+from pyspark.mllib.common import callMLlibFunc
 from pyspark.mllib.linalg import SparseVector, _convert_to_vector
 
 __all__ = ['LabeledPoint', 'LinearModel', 'LinearRegressionModel', 'RidgeRegressionModel',
@@ -129,8 +129,7 @@ def _regression_train_wrapper(train_func, modelClass, data, initial_weights):
     if not isinstance(first, LabeledPoint):
         raise ValueError("data should be an RDD of LabeledPoint, but got %s" % first)
     initial_weights = initial_weights or [0.0] * len(data.first().features)
-    weights, intercept = train_func(_to_java_object_rdd(data, cache=True),
-                                    _convert_to_vector(initial_weights))
+    weights, intercept = train_func(data, _convert_to_vector(initial_weights))
     return modelClass(weights, intercept)