19

I have a set of files. The path to the files are saved in a file., say all_files.txt. Using apache spark, I need to do an operation on all the files and club the results.

The steps that I want to do are:

  • Create an RDD by reading all_files.txt
  • For each line in all_files.txt (Each line is a path to some file), read the contents of each of the files into a single RDD
  • Then do an operation all contents

This is the code I wrote for the same:

def return_contents_from_file (file_name):
    return spark.read.text(file_name).rdd.map(lambda  r: r[0])

def run_spark():
    file_name = 'path_to_file'

    spark = SparkSession \
        .builder \
        .appName("PythonWordCount") \
        .getOrCreate()

    counts = spark.read.text(file_name).rdd.map(lambda r: r[0]) \ # this line is supposed to return the paths to each file
        .flatMap(return_contents_from_file) \ # here i am expecting to club all the contents of all files
        .flatMap(do_operation_on_each_line_of_all_files) # here i am expecting do an operation on each line of all files

This is throwing the error:

line 323, in get_return_value py4j.protocol.Py4JError: An error occurred while calling o25.getnewargs. Trace: py4j.Py4JException: Method getnewargs([]) does not exist at py4j.reflection.ReflectionEngine.getMethod(ReflectionEngine.java:318) at py4j.reflection.ReflectionEngine.getMethod(ReflectionEngine.java:326) at py4j.Gateway.invoke(Gateway.java:272) at py4j.commands.AbstractCommand.invokeMethod(AbstractCommand.java:132) at py4j.commands.CallCommand.execute(CallCommand.java:79) at py4j.GatewayConnection.run(GatewayConnection.java:214) at java.lang.Thread.run(Thread.java:745)

Can someone please tell me what I am doing wrong and how I should proceed further. Thanks in advance.

Nexaspx
  • 371
  • 4
  • 20
UnderWood
  • 803
  • 3
  • 12
  • 23

3 Answers3

22

Using spark inside flatMap or any transformation that occures on executors is not allowed (spark session is available on driver only). It is also not possible to create RDD of RDDs (see: Is it possible to create nested RDDs in Apache Spark?)

But you can achieve this transformation in another way - read all content of all_files.txt into dataframe, use local map to make them dataframes and local reduce to union all, see example:

>>> filenames = spark.read.text('all_files.txt').collect()
>>> dataframes = map(lambda r: spark.read.text(r[0]), filenames)
>>> all_lines_df = reduce(lambda df1, df2: df1.unionAll(df2), dataframes)
Community
  • 1
  • 1
Mariusz
  • 13,481
  • 3
  • 60
  • 64
  • Thank you for your reply. But how do I parallelize the whole process? wouldn't map(lambda r: spark.read.text(r[0]), filenames) serialize the whole process? – UnderWood Nov 07 '16 at 18:45
  • 1
    The process of reading files run in parallel, the only serialized part is building execution plan. Try it out! – Mariusz Nov 07 '16 at 19:09
6

I meet this problem today, finally figure out that I refered to a spark.DataFrame object in pandas_udf , which result to this error .

The conclution:

You can't use sparkSession object , spark.DataFrame object or other Spark distributed objects in udf and pandas_udf, because they are unpickled.

If you meet this error and you are using udf, check it carefully , must be relative problem.

Mithril
  • 12,947
  • 18
  • 102
  • 153
0

I also got this error trying to log my model with MLFlow using mlflow.sklearn.log_model when the model itself was a pyspark.ml.classification model. Using mlflow.spark.log_model solved the issue.

lucazav
  • 858
  • 9
  • 24