[SPARK-3491] [MLlib] [PySpark] use pickle to serialize data in MLlib

Currently, we serialize the data between JVM and Python case by case manually, this cannot scale to support so many APIs in MLlib.

This patch will try to address this problem by serialize the data using pickle protocol, using Pyrolite library to serialize/deserialize in JVM. Pickle protocol can be easily extended to support customized class.

All the modules are refactored to use this protocol.

Known issues: There will be some performance regression (both CPU and memory, the serialized data increased)

Author: Davies Liu <davies.liu@gmail.com>

Closes #2378 from davies/pickle_mllib and squashes the following commits:

dffbba2 [Davies Liu] Merge branch 'master' of github.com:apache/spark into pickle_mllib
810f97f [Davies Liu] fix equal of matrix
032cd62 [Davies Liu] add more type check and conversion for user_product
bd738ab [Davies Liu] address comments
e431377 [Davies Liu] fix cache of rdd, refactor
19d0967 [Davies Liu] refactor Picklers
2511e76 [Davies Liu] cleanup
1fccf1a [Davies Liu] address comments
a2cc855 [Davies Liu] fix tests
9ceff73 [Davies Liu] test size of serialized Rating
44e0551 [Davies Liu] fix cache
a379a81 [Davies Liu] fix pickle array in python2.7
df625c7 [Davies Liu] Merge commit '154d141' into pickle_mllib
154d141 [Davies Liu] fix autobatchedpickler
44736d7 [Davies Liu] speed up pickling array in Python 2.7
e1d1bfc [Davies Liu] refactor
708dc02 [Davies Liu] fix tests
9dcfb63 [Davies Liu] fix style
88034f0 [Davies Liu] rafactor, address comments
46a501e [Davies Liu] choose batch size automatically
df19464 [Davies Liu] memorize the module and class name during pickleing
f3506c5 [Davies Liu] Merge branch 'master' into pickle_mllib
722dd96 [Davies Liu] cleanup _common.py
0ee1525 [Davies Liu] remove outdated tests
b02e34f [Davies Liu] remove _common.py
84c721d [Davies Liu] Merge branch 'master' into pickle_mllib
4d7963e [Davies Liu] remove muanlly serialization
6d26b03 [Davies Liu] fix tests
c383544 [Davies Liu] classification
f2a0856 [Davies Liu] mllib/regression
d9f691f [Davies Liu] mllib/util
cccb8b1 [Davies Liu] mllib/tree
8fe166a [Davies Liu] Merge branch 'pickle' into pickle_mllib
aa2287e [Davies Liu] random
f1544c4 [Davies Liu] refactor clustering
52d1350 [Davies Liu] use new protocol in mllib/stat
b30ef35 [Davies Liu] use pickle to serialize data for mllib/recommendation
f44f771 [Davies Liu] enable tests about array
3908f5c [Davies Liu] Merge branch 'master' into pickle
c77c87b [Davies Liu] cleanup debugging code
60e4e2f [Davies Liu] support unpickle array.array for Python 2.6

1 files changed, 53 insertions, 16 deletions

diff --git a/python/pyspark/mllib/recommendation.py b/python/pyspark/mllib/recommendation.py
index 2df23394da..59c1c5ff0c 100644
--- a/python/pyspark/mllib/recommendation.py
+++ b/python/pyspark/mllib/recommendation.py
@@ -16,17 +16,25 @@
 #
 
 from pyspark import SparkContext
-from pyspark.mllib._common import \
-    _get_unmangled_rdd, _get_unmangled_double_vector_rdd, \
-    _serialize_double_matrix, _deserialize_double_matrix, \
-    _serialize_double_vector, _deserialize_double_vector, \
-    _get_initial_weights, _serialize_rating, _regression_train_wrapper, \
-    _serialize_tuple, RatingDeserializer
+from pyspark.serializers import PickleSerializer, AutoBatchedSerializer
 from pyspark.rdd import RDD
 
 __all__ = ['MatrixFactorizationModel', 'ALS']
 
 
+class Rating(object):
+    def __init__(self, user, product, rating):
+        self.user = int(user)
+        self.product = int(product)
+        self.rating = float(rating)
+
+    def __reduce__(self):
+        return Rating, (self.user, self.product, self.rating)
+
+    def __repr__(self):
+        return "Rating(%d, %d, %d)" % (self.user, self.product, self.rating)
+
+
 class MatrixFactorizationModel(object):
 
     """A matrix factorisation model trained by regularized alternating
@@ -39,7 +47,9 @@ class MatrixFactorizationModel(object):
     >>> model = ALS.trainImplicit(ratings, 1)
     >>> model.predict(2,2) is not None
     True
+
     >>> testset = sc.parallelize([(1, 2), (1, 1)])
+    >>> model = ALS.train(ratings, 1)
     >>> model.predictAll(testset).count() == 2
     True
     """
@@ -54,34 +64,61 @@ class MatrixFactorizationModel(object):
     def predict(self, user, product):
         return self._java_model.predict(user, product)
 
-    def predictAll(self, usersProducts):
-        usersProductsJRDD = _get_unmangled_rdd(usersProducts, _serialize_tuple)
-        return RDD(self._java_model.predict(usersProductsJRDD._jrdd),
-                   self._context, RatingDeserializer())
+    def predictAll(self, user_product):
+        assert isinstance(user_product, RDD), "user_product should be RDD of (user, product)"
+        first = user_product.first()
+        if isinstance(first, list):
+            user_product = user_product.map(tuple)
+            first = tuple(first)
+        assert type(first) is tuple and len(first) == 2, \
+            "user_product should be RDD of (user, product)"
+        if any(isinstance(x, str) for x in first):
+            user_product = user_product.map(lambda (u, p): (int(x), int(p)))
+            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())
+        jresult = self._java_model.predict(tuplerdd).toJavaRDD()
+        return RDD(sc._jvm.PythonRDD.javaToPython(jresult), sc,
+                   AutoBatchedSerializer(PickleSerializer()))
 
 
 class ALS(object):
 
     @classmethod
+    def _prepare(cls, ratings):
+        assert isinstance(ratings, RDD), "ratings should be RDD"
+        first = ratings.first()
+        if not isinstance(first, Rating):
+            if isinstance(first, (tuple, list)):
+                ratings = ratings.map(lambda x: Rating(*x))
+            else:
+                raise ValueError("rating should be RDD of Rating or tuple/list")
+        # 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()
+
+    @classmethod
     def train(cls, ratings, rank, iterations=5, lambda_=0.01, blocks=-1):
         sc = ratings.context
-        ratingBytes = _get_unmangled_rdd(ratings, _serialize_rating)
-        mod = sc._jvm.PythonMLLibAPI().trainALSModel(
-            ratingBytes._jrdd, rank, iterations, lambda_, blocks)
+        jrating = cls._prepare(ratings)
+        mod = sc._jvm.PythonMLLibAPI().trainALSModel(jrating, rank, iterations, lambda_, blocks)
         return MatrixFactorizationModel(sc, mod)
 
     @classmethod
     def trainImplicit(cls, ratings, rank, iterations=5, lambda_=0.01, blocks=-1, alpha=0.01):
         sc = ratings.context
-        ratingBytes = _get_unmangled_rdd(ratings, _serialize_rating)
+        jrating = cls._prepare(ratings)
         mod = sc._jvm.PythonMLLibAPI().trainImplicitALSModel(
-            ratingBytes._jrdd, rank, iterations, lambda_, blocks, alpha)
+            jrating, rank, iterations, lambda_, blocks, alpha)
         return MatrixFactorizationModel(sc, mod)
 
 
 def _test():
     import doctest
-    globs = globals().copy()
+    import pyspark.mllib.recommendation
+    globs = pyspark.mllib.recommendation.__dict__.copy()
     globs['sc'] = SparkContext('local[4]', 'PythonTest', batchSize=2)
     (failure_count, test_count) = doctest.testmod(globs=globs, optionflags=doctest.ELLIPSIS)
     globs['sc'].stop()