PySpark: Convert JSON String Column to Array of Object (StructType) in Data Frame

This post shows how to derive new column in a Spark data frame from a JSON array string column. I am running the code in Spark 2.2.1 though it is compatible with Spark 1.6.0 (with less JSON SQL functions).

Prerequisites

Refer to the following post to install Spark in Windows.

Install Spark 2.2.1 in Windows

*If you are using Linux or UNIX, the code should also work.

Requirement

Convert the following list to a data frame:

source = [{"attr_1": 1, "attr_2": "[{\"a\":1,\"b\":1},{\"a\":2,\"b\":2}]"}, {
"attr_1": 2, "attr_2": "[{\"a\":3,\"b\":3},{\"a\":4,\"b\":4}]"}]

The data frame should have two column:

• attr_1: column type is IntegerType
• attr_2: column type is ArrayType (element type is StructType with two StructField).

And the schema of the data frame should look like the following:

root
  |-- attr_1: long (nullable = true)
  |-- attr_2: array (nullable = true)
  |    |-- element: struct (containsNull = true)
  |    |    |-- a: integer (nullable = false)
  |    |    |-- b: integer (nullable = false)

Resolution

Convert list to data frame

First, let’s convert the list to a data frame in Spark by using the following code:

# Read the list into data frame
df = sqlContext.read.json(sc.parallelize(source))
df.show()
df.printSchema()

JSON is read into a data frame through sqlContext. The output is:

+------+--------------------+
|attr_1|              attr_2|
+------+--------------------+
|     1|[{"a":1,"b":1},{"...|
|     2|[{"a":3,"b":3},{"...|
+------+--------------------+

root
  |-- attr_1: long (nullable = true)
  |-- attr_2: string (nullable = true)

At current stage, column attr_2 is string type instead of array of struct.

Create a function to parse JSON to list

For column attr_2, the value is JSON array string. Let’s create a function to parse JSON string and then convert it to list.

# Function to convert JSON array string to a list
import json
def parse_json(array_str):
json_obj = json.loads(array_str)
for item in json_obj:
   yield (item["a"], item["b"])

Define the schema of column attr_2

# Define the schema
from pyspark.sql.types import ArrayType, IntegerType, StructType, StructField
json_schema = ArrayType(StructType([StructField('a', IntegerType(
), nullable=False), StructField('b', IntegerType(), nullable=False)]))

Based on the JSON string, the schema is defined as an array of struct with two fields.

Create an UDF

Now, we can create an UDF with function parse_json and schema json_schema.

# Define udf
from pyspark.sql.functions import udf
udf_parse_json = udf(lambda str: parse_json(str), json_schema)

Create a new data frame

Finally, we can create a new data frame using the defined UDF.

# Generate a new data frame with the expected schema
df_new = df.select(df.attr_1, udf_parse_json(df.attr_2).alias("attr_2"))
df_new.show()
df_new.printSchema()

The output is as the following:

+------+--------------+
|attr_1|        attr_2|
+------+--------------+
|     1|[[1,1], [2,2]]|
|     2|[[3,3], [4,4]]|
+------+--------------+

root
  |-- attr_1: long (nullable = true)
  |-- attr_2: array (nullable = true)
  |    |-- element: struct (containsNull = true)
  |    |    |-- a: integer (nullable = false)
  |    |    |-- b: integer (nullable = false)

Summary

The following is the complete code:

from pyspark import SparkContext, SparkConf, SQLContext
appName = "JSON Parse Example"
master = "local[2]"
conf = SparkConf().setAppName(appName).setMaster(master)
sc = SparkContext(conf=conf)
sqlContext = SQLContext(sc)
source = [{"attr_1": 1, "attr_2": "[{\"a\":1,\"b\":1},{\"a\":2,\"b\":2}]"}, {
"attr_1": 2, "attr_2": "[{\"a\":3,\"b\":3},{\"a\":4,\"b\":4}]"}]
# Read the list into data frame
df = sqlContext.read.json(sc.parallelize(source))
df.show()
df.printSchema()
# Function to convert JSON array string to a list
import json
def parse_json(array_str):
json_obj = json.loads(array_str)
for item in json_obj:
yield (item["a"], item["b"])
# Define the schema
from pyspark.sql.types import ArrayType, IntegerType, StructType, StructField
json_schema = ArrayType(StructType([StructField('a', IntegerType(
), nullable=False), StructField('b', IntegerType(), nullable=False)]))
# Define udf
from pyspark.sql.functions import udf
udf_parse_json = udf(lambda str: parse_json(str), json_schema)
# Generate a new data frame with the expected schema
df_new = df.select(df.attr_1, udf_parse_json(df.attr_2).alias("attr_2"))
df_new.show()
df_new.printSchema()

Save the code as file parse_json.py and then you can use the following command to run it in Spark:

spark-submit parse_json.py

The following screenshot is captured from my local environment (Spark 2.2.1 & Python 3.6.4 in Windows ).