Convert a JSON string to a struct column without schema in Spark

Question

Spark: 3.0.0
Scala: 2.12.8

My data frame has a column with JSON string, and I want to create a new column from it with the StructType.

temp_json_string
{"name":"test","id":"12","category":[{"products":["A","B"],"displayName":"test_1","displayLabel":"test1"},{"products":["C"],"displayName":"test_2","displayLabel":"test2"}],"createdAt":"","createdBy":""}

root
 |-- temp_json_string: string (nullable = true)

Formatted JSON:

{
  "name":"test",
  "id":"12",
  "category":[
    {
      "products":[
        "A",
        "B"
      ],
      "displayName":"test_1",
      "displayLabel":"test1"
    },
    {
      "products":[
        "C"
      ],
      "displayName":"test_2",
      "displayLabel":"test2"
    }
  ],
  "createdAt":"",
  "createdBy":""
}

I want to create a new column of type Struct so I tried:

dataFrame
     .withColumn("temp_json_struct", struct(col("temp_json_string")))
     .select("temp_json_struct")

Now, I get the schema as:

root
 |-- temp_json_struct: struct (nullable = false)
 |    |-- temp_json_string: string (nullable = true)

Desired result:

root
 |-- temp_json_struct: struct (nullable = false)
 |    |-- name: string (nullable = true)
 |    |-- category: array (nullable = true)
 |    |    |-- products: array (nullable = true)
 |    |    |-- displayName: string (nullable = true)
 |    |    |-- displayLabel: string (nullable = true)
 |    |-- createdAt: timestamp (nullable = true)
 |    |-- updatedAt: timestamp (nullable = true)

Answer 1

If your trying to quickly debug or dig into data, a simple SQL solution you can pretty easily implement is to work out the schema yourself.

e.g.

strJsonContent
--------------
[{"a":1,"b":"xyz"}, {"a":2,"b":"xyz"}]

Schema would be:

ARRAY<STRUCT<`a`: BIGINT, `b`: STRING>>

Want to put this into a query:

select 
  from_json(strJsonContent, 'ARRAY<STRUCT<`a`: BIGINT, `b`: STRING>>')) as structured_strJsonContent, *
  from tableName

Answer 2

json_str_col is the column that has JSON string. I had multiple files so that's why the fist line is iterating through each row to extract the schema. If you know your schema up front then just replace json_schema with that.

json_schema = spark.read.json(df.rdd.map(lambda row: row.json_str_col)).schema
df = df.withColumn('new_col', from_json(col('json_str_col'), json_schema))

Answer 3

// import spark implicits for conversion to dataset (.as[String])
import spark.implicits._

val df = ??? //create your dataframe having the 'temp_json_string' column

//convert Dataset[Row] aka Dataframe to Dataset[String]
val ds = df.select("temp_json_string").as[String]

//read as json
spark.read.json(ds)

Documentation

Answer 4

There at least two different ways to retrieve/discover the schema for a given JSON.

For the illustration, let's create some data first:

import org.apache.spark.sql.types.StructType

val jsData = Seq(
  ("""{
    "name":"test","id":"12","category":[
    {
      "products":[
        "A",
        "B"
      ],
      "displayName":"test_1",
      "displayLabel":"test1"
    },
    {
      "products":[
        "C"
      ],
      "displayName":"test_2",
      "displayLabel":"test2"
    }
  ],
  "createdAt":"",
  "createdBy":""}""")
)

Option 1: schema_of_json

The first option is to use the built-in function schema_of_json. The function will return the schema for the given JSON in DDL format:

val json = jsData.toDF("js").collect()(0).getString(0)

val ddlSchema: String = spark.sql(s"select schema_of_json('${json}')")
                            .collect()(0) //get 1st row
                            .getString(0) //get 1st col of the row as string
                            .replace("null", "string") //replace type with string, this occurs since you have "createdAt":"" 

// struct<category:array<struct<displayLabel:string,displayName:string,products:array<string>>>,createdAt:null,createdBy:null,id:string,name:string>

val schema: StructType = StructType.fromDDL(s"js_schema $ddlSchema")

Note that you would expect that schema_of_json would also work on the column level i.e: schema_of_json(js_col), unfortunately, this doesn't work as expected therefore we are forced to pass a string instead.

Option 2: use Spark JSON reader (recommended)

import org.apache.spark.sql.functions.from_json

val schema: StructType = spark.read.json(jsData.toDS).schema

// schema.printTreeString

// root
//  |-- category: array (nullable = true)
//  |    |-- element: struct (containsNull = true)
//  |    |    |-- displayLabel: string (nullable = true)
//  |    |    |-- displayName: string (nullable = true)
//  |    |    |-- products: array (nullable = true)
//  |    |    |    |-- element: string (containsNull = true)
//  |-- createdAt: string (nullable = true)
//  |-- createdBy: string (nullable = true)
//  |-- id: string (nullable = true)
//  |-- name: string (nullable = true)

As you can see, here we are producing a schema based on StructType and not a DDL string as in the previous case.

After discovering the schema we can move on to the next step which is converting the JSON data into a struct. To achieve that we will use from_json built-in function:

jsData.toDF("js")
      .withColumn("temp_json_struct", from_json($"js", schema))
      .printSchema()

// root
//  |-- js: string (nullable = true)
//  |-- temp_json_struct: struct (nullable = true)
//  |    |-- category: array (nullable = true)
//  |    |    |-- element: struct (containsNull = true)
//  |    |    |    |-- displayLabel: string (nullable = true)
//  |    |    |    |-- displayName: string (nullable = true)
//  |    |    |    |-- products: array (nullable = true)
//  |    |    |    |    |-- element: string (containsNull = true)
//  |    |-- createdAt: string (nullable = true)
//  |    |-- createdBy: string (nullable = true)
//  |    |-- id: string (nullable = true)
//  |    |-- name: string (nullable = true)