Databricks - Pyspark - Handling nested json with a dynamic key

Question

I have an example json data file which has the following structure:

{
    "Header": {
        "Code1": "abc",
        "Code2": "def",
        "Code3": "ghi",
        "Code4": "jkl",
    },
    "TimeSeries": {
        "2020-11-25T03:00:00+00:00": {
            "UnitPrice": 1000,
            "Amount": 10000,

        },
        "2020-11-26T03:00:00+00:00": {
            "UnitPrice": 1000,
            "Amount": 10000,

        }
    }
}

When I parse this into databricks with command:

df = spark.read.json("/FileStore/test.txt")

I get as output 2 objects: Header and TimeSeries. With the TimeSeries I want to be able to flatten the structure so it has the following schema:

Date
UnitPrice
Amount

As the date field is a key, I am currently only able to access it via iterating through the column names and then using this in the dot-notation dynamically:

def flatten_json(data):


  columnlist = data.select("TimeSeries.*")
  count = 0 
  for name in data.select("TimeSeries.*"):
    df1 = data.select("Header.*").withColumn(("Timeseries"), lit(columnlist.columns[count])).withColumn("join", lit("a"))
    df2 = data.select("TimeSeries." + columnlist.columns[count] + ".*").withColumn("join", lit("a"))
    if count == 0: 
      df3 = df1.join(df2, on=['join'], how="inner")
    else: 
      df3 = df3.union(df1.join(df2, on=['join'], how="inner"))
    count = count + 1
  return(df3)

This is far from ideal. Does anyone know a better method to create the described dataframe?

user_3546203 · Answer

The idea:

Step 1: Extract Header and TimeSeries separately.
Step 2: For each field in the TimeSeries object, extract the Amount and UnitPrice, together with the name of the field, stuff them into a struct.
Step 3: Merge all these structs into an array column, and explode it.
Step 4: Extract Timeseries, Amount and UnitPrice from the exploded column.
Step 5: Cross join with the Header row.

import pyspark.sql.functions as F

header_df = df.select("Header.*")
timeseries_df = df.select("TimeSeries.*")
fieldNames = enumerate(timeseries_df.schema.fieldNames())
cols = [F.struct(F.lit(name).alias("Timeseries"), col(name).getItem("Amount").alias("Amount"), col(name).getItem("UnitPrice").alias("UnitPrice")).alias("ts_" + str(idx)) for idx, name in fieldNames]
combined = explode(array(cols)).alias("comb")
timeseries = timeseries_df.select(combined).select('comb.Timeseries', 'comb.Amount', 'comb.UnitPrice')
result = header_df.crossJoin(timeseries)
result.show(truncate = False)

Output:

+-----+-----+-----+-----+-------------------------+------+---------+
|Code1|Code2|Code3|Code4|Timeseries               |Amount|UnitPrice|
+-----+-----+-----+-----+-------------------------+------+---------+
|abc  |def  |ghi  |jkl  |2020-11-25T03:00:00+00:00|10000 |1000     |
|abc  |def  |ghi  |jkl  |2020-11-26T03:00:00+00:00|10000 |1000     |
+-----+-----+-----+-----+-------------------------+------+---------+

Databricks - Pyspark - Handling nested json with a dynamic key

1 Answers