'PipelinedRDD' object has no attribute 'toDF' in PySpark

Question

I'm trying to load an SVM file and convert it to a DataFrame so I can use the ML module (Pipeline ML) from Spark. I've just installed a fresh Spark 1.5.0 on an Ubuntu 14.04 (no spark-env.sh configured).

My my_script.py is:

from pyspark.mllib.util import MLUtils
from pyspark import SparkContext

sc = SparkContext("local", "Teste Original")
data = MLUtils.loadLibSVMFile(sc, "/home/svm_capture").toDF()

and I'm running using: ./spark-submit my_script.py

And I get the error:

Traceback (most recent call last):
File "/home/fred-spark/spark-1.5.0-bin-hadoop2.6/pipeline_teste_original.py", line 34, in <module>
data = MLUtils.loadLibSVMFile(sc, "/home/fred-spark/svm_capture").toDF()
AttributeError: 'PipelinedRDD' object has no attribute 'toDF'

What I can't understand is that if I run:

data = MLUtils.loadLibSVMFile(sc, "/home/svm_capture").toDF()

directly inside PySpark shell, it works.

Solution 1

toDF method is a monkey patch executed inside SparkSession (SQLContext constructor in 1.x) constructor so to be able to use it you have to create a SQLContext (or SparkSession) first:

# SQLContext or HiveContext in Spark 1.x
from pyspark.sql import SparkSession
from pyspark import SparkContext

sc = SparkContext()

rdd = sc.parallelize([("a", 1)])
hasattr(rdd, "toDF")
## False

spark = SparkSession(sc)
hasattr(rdd, "toDF")
## True

rdd.toDF().show()
## +---+---+
## | _1| _2|
## +---+---+
## |  a|  1|
## +---+---+

Not to mention you need a SQLContext or SparkSession to work with DataFrames in the first place.

Solution 2

Make sure you have spark session too.

sc = SparkContext("local", "first app")
spark = SparkSession(sc)