Passing a data frame column and external list to udf under withColumn

Question

I have a Spark dataframe with the following structure. The bodyText_token has the tokens (processed/set of words). And I have a nested list of defined keywords

root
 |-- id: string (nullable = true)
 |-- body: string (nullable = true)
 |-- bodyText_token: array (nullable = true)

keyword_list=[['union','workers','strike','pay','rally','free','immigration',],
['farmer','plants','fruits','workers'],['outside','field','party','clothes','fashions']]

I needed to check how many tokens fall under each keyword list and add the result as a new column of the existing dataframe. Eg: if tokens =["become", "farmer","rally","workers","student"] the result will be -> [1,2,0]

The following function worked as expected.

def label_maker_topic(tokens,topic_words):
    twt_list = []
    for i in range(0, len(topic_words)):
        count = 0
        #print(topic_words[i])
        for tkn in tokens:
            if tkn in topic_words[i]:
                count += 1
        twt_list.append(count)
    
    return twt_list

I used udf under withColumn to access the function and I get an error. I think it's about passing an external list to a udf. Is there a way I can pass the external list and the dataframe column to a udf and add a new column to my dataframe?

topicWord = udf(label_maker_topic,StringType())
myDF=myDF.withColumn("topic_word_count",topicWord(myDF.bodyText_token,keyword_list))

Solution 1

The cleanest solution is to pass additional arguments using closure:

def make_topic_word(topic_words):
     return udf(lambda c: label_maker_topic(c, topic_words))

df = sc.parallelize([(["union"], )]).toDF(["tokens"])

(df.withColumn("topics", make_topic_word(keyword_list)(col("tokens")))
    .show())

This doesn't require any changes in keyword_list or the function you wrap with UDF. You can also use this method to pass an arbitrary object. This can be used to pass for example a list of sets for efficient lookups.

If you want to use your current UDF and pass topic_words directly you'll have to convert it to a column literal first:

from pyspark.sql.functions import array, lit

ks_lit = array(*[array(*[lit(k) for k in ks]) for ks in keyword_list])
df.withColumn("ad", topicWord(col("tokens"), ks_lit)).show()

Depending on your data and requirements there can alternative, more efficient solutions, which don't require UDFs (explode + aggregate + collapse) or lookups (hashing + vector operations).

Solution 2

The following works fine where any external parameter can be passed to the UDF (a tweaked code to help anyone)

topicWord=udf(lambda tkn: label_maker_topic(tkn,topic_words),StringType())
myDF=myDF.withColumn("topic_word_count",topicWord(myDF.bodyText_token))

Solution 3

The keyword_list list should be broadcasted to all the nodes in the cluster if the list is big. I'm guessing zero's solution works because the list is tiny and is auto-broadcasted. It's better to explicitly broadcast in my opinion to leave no doubts (explicitly broadcasting is required for bigger lists).

keyword_list=[
    ['union','workers','strike','pay','rally','free','immigration',],
    ['farmer','plants','fruits','workers'],
    ['outside','field','party','clothes','fashions']]

def label_maker_topic(tokens, topic_words_broadcasted):
    twt_list = []
    for i in range(0, len(topic_words_broadcasted.value)):
        count = 0
        #print(topic_words[i])
        for tkn in tokens:
            if tkn in topic_words_broadcasted.value[i]:
                count += 1
        twt_list.append(count)

    return twt_list

def make_topic_word_better(topic_words_broadcasted):
    def f(c):
        return label_maker_topic(c, topic_words_broadcasted)
    return F.udf(f)

df = spark.createDataFrame([["union",], ["party",]]).toDF("tokens")
b = spark.sparkContext.broadcast(keyword_list)
df.withColumn("topics", make_topic_word_better(b)(F.col("tokens"))).show()

Here's what'll be outputted:

+------+---------+
|tokens|   topics|
+------+---------+
| union|[0, 0, 0]|
| party|[0, 0, 0]|
+------+---------+

Note that you need to call value to access the list that's been broadcasted (e.g. topic_words_broadcasted.value). It's a difficult implementation, but important to master because a lot of PySpark UDFs rely on a list or dictionary that's been broadcasted.

Solution 4

Just the other way using partial from functools module

from functools import partial

func_to_call = partial(label_maker_topic, topic_words=keyword_list)

pyspark_udf = udf(func_to_call, <specify_the_type_returned_by_function_here>)

df = sc.parallelize([(["union"], )]).toDF(["tokens"])

df.withColumn("topics", pyspark_udf(col("tokens"))).show()

View more solutions