Implement a java UDF and call it from pyspark

问题:

I need to create a UDF to be used in pyspark python which uses a java object for its internal calculations.

If it were a simple python I would do something like:

def f(x):
    return 7
fudf = pyspark.sql.functions.udf(f,pyspark.sql.types.IntegerType())

and call it using:

df = sqlContext.range(0,5)
df2 = df.withColumn("a",fudf(df.id)).show()

However, the implementation of the function I need is in java and not in python. I need to wrap it somehow so I can call it in a similar way from python.

My first try was to do implement the java object, then wrap it in python in pyspark and convert that to UDF. That failed with serialization error.

Java code:

package com.test1.test2;

public class TestClass1 {
    Integer internalVal;
    public TestClass1(Integer val1) {
        internalVal = val1;
    }
    public Integer do_something(Integer val) {
        return internalVal;
    }    
}

pyspark code:

from py4j.java_gateway import java_import
from pyspark.sql.functions import udf
from pyspark.sql.types import IntegerType
java_import(sc._gateway.jvm, "com.test1.test2.TestClass1")
a = sc._gateway.jvm.com.test1.test2.TestClass1(7)
audf = udf(a,IntegerType())

error:

---------------------------------------------------------------------------
Py4JError                                 Traceback (most recent call last)
<ipython-input-2-9756772ab14f> in <module>()
      4 java_import(sc._gateway.jvm, "com.test1.test2.TestClass1")
      5 a = sc._gateway.jvm.com.test1.test2.TestClass1(7)
----> 6 audf = udf(a,IntegerType())

/usr/local/spark/python/pyspark/sql/functions.py in udf(f, returnType)
   1595     [Row(slen=5), Row(slen=3)]
   1596     """
-> 1597     return UserDefinedFunction(f, returnType)
   1598 
   1599 blacklist = ['map', 'since', 'ignore_unicode_prefix']

/usr/local/spark/python/pyspark/sql/functions.py in __init__(self, func, returnType, name)
   1556         self.returnType = returnType
   1557         self._broadcast = None
-> 1558         self._judf = self._create_judf(name)
   1559 
   1560     def _create_judf(self, name):

/usr/local/spark/python/pyspark/sql/functions.py in _create_judf(self, name)
   1565         command = (func, None, ser, ser)
   1566         sc = SparkContext.getOrCreate()
-> 1567         pickled_command, broadcast_vars, env, includes = _prepare_for_python_RDD(sc, command, self)
   1568         ctx = SQLContext.getOrCreate(sc)
   1569         jdt = ctx._ssql_ctx.parseDataType(self.returnType.json())

/usr/local/spark/python/pyspark/rdd.py in _prepare_for_python_RDD(sc, command, obj)
   2297     # the serialized command will be compressed by broadcast
   2298     ser = CloudPickleSerializer()
-> 2299     pickled_command = ser.dumps(command)
   2300     if len(pickled_command) > (1 << 20):  # 1M
   2301         # The broadcast will have same life cycle as created PythonRDD

/usr/local/spark/python/pyspark/serializers.py in dumps(self, obj)
    426 
    427     def dumps(self, obj):
--> 428         return cloudpickle.dumps(obj, 2)
    429 
    430 

/usr/local/spark/python/pyspark/cloudpickle.py in dumps(obj, protocol)
    644 
    645     cp = CloudPickler(file,protocol)
--> 646     cp.dump(obj)
    647 
    648     return file.getvalue()

/usr/local/spark/python/pyspark/cloudpickle.py in dump(self, obj)
    105         self.inject_addons()
    106         try:
--> 107             return Pickler.dump(self, obj)
    108         except RuntimeError as e:
    109             if 'recursion' in e.args[0]:

/home/mendea3/anaconda2/lib/python2.7/pickle.pyc in dump(self, obj)
    222         if self.proto >= 2:
    223             self.write(PROTO + chr(self.proto))
--> 224         self.save(obj)
    225         self.write(STOP)
    226 

/home/mendea3/anaconda2/lib/python2.7/pickle.pyc in save(self, obj)
    284         f = self.dispatch.get(t)
    285         if f:
--> 286             f(self, obj) # Call unbound method with explicit self
    287             return
    288 

/home/mendea3/anaconda2/lib/python2.7/pickle.pyc in save_tuple(self, obj)
    566         write(MARK)
    567         for element in obj:
--> 568             save(element)
    569 
    570         if id(obj) in memo:

/home/mendea3/anaconda2/lib/python2.7/pickle.pyc in save(self, obj)
    284         f = self.dispatch.get(t)
    285         if f:
--> 286             f(self, obj) # Call unbound method with explicit self
    287             return
    288 

/usr/local/spark/python/pyspark/cloudpickle.py in save_function(self, obj, name)
    191         if islambda(obj) or obj.__code__.co_filename == '<stdin>' or themodule is None:
    192             #print("save global", islambda(obj), obj.__code__.co_filename, modname, themodule)
--> 193             self.save_function_tuple(obj)
    194             return
    195         else:

/usr/local/spark/python/pyspark/cloudpickle.py in save_function_tuple(self, func)
    234         # create a skeleton function object and memoize it
    235         save(_make_skel_func)
--> 236         save((code, closure, base_globals))
    237         write(pickle.REDUCE)
    238         self.memoize(func)

/home/mendea3/anaconda2/lib/python2.7/pickle.pyc in save(self, obj)
    284         f = self.dispatch.get(t)
    285         if f:
--> 286             f(self, obj) # Call unbound method with explicit self
    287             return
    288 

/home/mendea3/anaconda2/lib/python2.7/pickle.pyc in save_tuple(self, obj)
    552         if n <= 3 and proto >= 2:
    553             for element in obj:
--> 554                 save(element)
    555             # Subtle.  Same as in the big comment below.
    556             if id(obj) in memo:

/home/mendea3/anaconda2/lib/python2.7/pickle.pyc in save(self, obj)
    284         f = self.dispatch.get(t)
    285         if f:
--> 286             f(self, obj) # Call unbound method with explicit self
    287             return
    288 

/home/mendea3/anaconda2/lib/python2.7/pickle.pyc in save_list(self, obj)
    604 
    605         self.memoize(obj)
--> 606         self._batch_appends(iter(obj))
    607 
    608     dispatch[ListType] = save_list

/home/mendea3/anaconda2/lib/python2.7/pickle.pyc in _batch_appends(self, items)
    637                 write(MARK)
    638                 for x in tmp:
--> 639                     save(x)
    640                 write(APPENDS)
    641             elif n:

/home/mendea3/anaconda2/lib/python2.7/pickle.pyc in save(self, obj)
    304             reduce = getattr(obj, "__reduce_ex__", None)
    305             if reduce:
--> 306                 rv = reduce(self.proto)
    307             else:
    308                 reduce = getattr(obj, "__reduce__", None)

/usr/local/spark/python/lib/py4j-0.9-src.zip/py4j/java_gateway.py in __call__(self, *args)
    811         answer = self.gateway_client.send_command(command)
    812         return_value = get_return_value(
--> 813             answer, self.gateway_client, self.target_id, self.name)
    814 
    815         for temp_arg in temp_args:

/usr/local/spark/python/pyspark/sql/utils.py in deco(*a, **kw)
     43     def deco(*a, **kw):
     44         try:
---> 45             return f(*a, **kw)
     46         except py4j.protocol.Py4JJavaError as e:
     47             s = e.java_exception.toString()

/usr/local/spark/python/lib/py4j-0.9-src.zip/py4j/protocol.py in get_return_value(answer, gateway_client, target_id, name)
    310                 raise Py4JError(
    311                     "An error occurred while calling {0}{1}{2}. Trace:\n{3}\n".
--> 312                     format(target_id, ".", name, value))
    313         else:
    314             raise Py4JError(

Py4JError: An error occurred while calling o18.__getnewargs__. Trace:
py4j.Py4JException: Method __getnewargs__([]) does not exist
    at py4j.reflection.ReflectionEngine.getMethod(ReflectionEngine.java:335)
    at py4j.reflection.ReflectionEngine.getMethod(ReflectionEngine.java:344)
    at py4j.Gateway.invoke(Gateway.java:252)
    at py4j.commands.AbstractCommand.invokeMethod(AbstractCommand.java:133)
    at py4j.commands.CallCommand.execute(CallCommand.java:79)
    at py4j.GatewayConnection.run(GatewayConnection.java:209)
    at java.lang.Thread.run(Thread.java:745)

EDIT: I also tried to make the java class serializable but to no avail.

My second attempt was to define the UDF in java to begin with but that failed as I am not sure how to correctly wrap it:

java code: package com.test1.test2;

import org.apache.spark.sql.api.java.UDF1;

public class TestClassUdf implements UDF1<Integer, Integer> {

    Integer retval;

    public TestClassUdf(Integer val) {
        retval = val;
    }

    @Override
    public Integer call(Integer arg0) throws Exception {
        return retval;
    }   
}

but how would I use it? I tried:

from py4j.java_gateway import java_import
java_import(sc._gateway.jvm, "com.test1.test2.TestClassUdf")
a = sc._gateway.jvm.com.test1.test2.TestClassUdf(7)
dfint = sqlContext.range(0,15)
df = dfint.withColumn("a",a(dfint.id))

but I get:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-5-514811090b5f> in <module>()
      3 a = sc._gateway.jvm.com.test1.test2.TestClassUdf(7)
      4 dfint = sqlContext.range(0,15)
----> 5 df = dfint.withColumn("a",a(dfint.id))

TypeError: 'JavaObject' object is not callable

and I tried to use a.call instead of a:

df = dfint.withColumn("a",a.call(dfint.id))

but got: --------------------------------------------------------------------------- TypeError Traceback (most recent call last) in () 3 a = sc._gateway.jvm.com.test1.test2.TestClassUdf(7) 4 dfint = sqlContext.range(0,15) ----> 5 df = dfint.withColumn("a",a.call(dfint.id))

/usr/local/spark/python/lib/py4j-0.9-src.zip/py4j/java_gateway.py in __call__(self, *args)
    796     def __call__(self, *args):
    797         if self.converters is not None and len(self.converters) > 0:
--> 798             (new_args, temp_args) = self._get_args(args)
    799         else:
    800             new_args = args

/usr/local/spark/python/lib/py4j-0.9-src.zip/py4j/java_gateway.py in _get_args(self, args)
    783                 for converter in self.gateway_client.converters:
    784                     if converter.can_convert(arg):
--> 785                         temp_arg = converter.convert(arg, self.gateway_client)
    786                         temp_args.append(temp_arg)
    787                         new_args.append(temp_arg)

/usr/local/spark/python/lib/py4j-0.9-src.zip/py4j/java_collections.py in convert(self, object, gateway_client)
    510         HashMap = JavaClass("java.util.HashMap", gateway_client)
    511         java_map = HashMap()
--> 512         for key in object.keys():
    513             java_map[key] = object[key]
    514         return java_map

TypeError: 'Column' object is not callable

Any help would be appriciated.

回答1:

I got this working with the help of another question (and answer) of your own about UDAFs.

Spark provides a udf() method for wrapping Scala FunctionN, so we can wrap the Java function in Scala and use that. Your Java method needs to be static or on a class that implements Serializable.

package com.example

import org.apache.spark.sql.UserDefinedFunction
import org.apache.spark.sql.functions.udf

class MyUdf extends Serializable {
  def getUdf: UserDefinedFunction = udf(() => MyJavaClass.MyJavaMethod())
}

Usage in PySpark:

def my_udf():
    from pyspark.sql.column import Column, _to_java_column, _to_seq
    pcls = "com.example.MyUdf"
    jc = sc._jvm.java.lang.Thread.currentThread() \
        .getContextClassLoader().loadClass(pcls).newInstance().getUdf().apply
    return Column(jc(_to_seq(sc, [], _to_java_column)))

rdd1 = sc.parallelize([{'c1': 'a'}, {'c1': 'b'}, {'c1': 'c'}])
df1 = rdd1.toDF()
df2 = df1.withColumn('mycol', my_udf())

As with the UDAF in your other question and answer, we can pass columns into it with return Column(jc(_to_seq(sc, ["col1", "col2"], _to_java_column)))

回答2:

In lines with https://dzone.com/articles/pyspark-java-udf-integration-1 you could define UDF1 with in Java using

public class AddNumber implements UDF1<Long, Long> {

@Override
public Long call(Long num) throws Exception {
      return (num + 5);
   }
}

And then after adding the jar to your pyspark with --package <your-jar>

you can use it in pyspark as:

from pyspark.sql import functions as F
from pyspark.sql.types import LongType


>>> df = spark.createDataFrame([float(i) for i in range(100)], FloatType()).toDF("a")
>>> spark.udf.registerJavaFunction("addNumber", "com.example.spark.AddNumber", LongType())
>>> df.withColumn("b", F.expr("addNumber(a)")).show(5)
+---+---+
|  a|  b|
+---+---+
|0.0|  5|
|1.0|  6|
|2.0|  7|
|3.0|  8|
|4.0|  8|
+---+---+
only showing top 5 rows