flatMap over list of custom objects in pyspark

问题

I'm getting an error when running flatMap() on a list of objects of a class. It works fine for regular python data types like int, list etc. but I'm facing an error when the list contains objects of my class. Here's the entire code:

from pyspark import SparkContext 

sc = SparkContext("local","WordCountBySparkKeyword")

def func(x):
    if x==2:
        return [2, 3, 4]
    return [1]

rdd = sc.parallelize([2])
rdd = rdd.flatMap(func) # rdd.collect() now has [2, 3, 4]
rdd = rdd.flatMap(func) # rdd.collect() now has [2, 3, 4, 1, 1]

print rdd.collect() # gives expected output

# Class I'm defining
class node(object):
    def __init__(self, value):
        self.value = value

    # Representation, for printing node
    def __repr__(self):
        return self.value


def foo(x):
    if x.value==2:
        return [node(2), node(3), node(4)]
    return [node(1)]

rdd = sc.parallelize([node(2)])
rdd = rdd.flatMap(foo)  #marker 2

print rdd.collect() # rdd.collect should contain nodes with values [2, 3, 4, 1, 1]

The code works fine till marker 1(commented in code). The problem arises after marker 2. The specific error message I'm getting is AttributeError: 'module' object has no attribute 'node' How do I resolve this error?

I'm working on ubuntu, running pyspark 1.4.1

回答1:

Error you get is completely unrelated to flatMap. If you define node class in your main script it is accessible on a driver but it is not distributed to the workers. To make it work you should place node definition inside separate module and makes sure it is distributed to the workers.

Create separate module with node definition, lets call it node.py
Import this node class inside your main script:
```
from node import node
```
Make sure module is distributed to the workers:
```
sc.addPyFile("node.py")
```

Now everything should work as expected.

On a side note:

PEP 8 recommends CapWords for class names. It is not a hard requirement but it makes life easier
__repr__ method should return a string representation of an object. At least make sure it is a string, but a proper representation is even better:
```
def __repr__(self):
     return "node({0})".format(repr(self.value))
```

来源：https://stackoverflow.com/questions/32792271/flatmap-over-list-of-custom-objects-in-pyspark

标签

python

apache-spark

pyspark

flatmap