Spark structured streaming: Schema Inference in Scala

Question

I'm trying to infer the dynamic json schema from kafka topic.Found this piece of code in blog, which infer the schema using PYSPARK.

  def read_kafka_topic(topic):
    
    df_json = (spark.read
               .format("kafka")
               .option("kafka.bootstrap.servers", kafka_broker)
               .option("subscribe", topic)
               .option("startingOffsets", "earliest")
               .option("endingOffsets", "latest")
               .option("failOnDataLoss", "false")
               .load()
               .withColumn("value", expr("string(value)"))
               .filter(col("value").isNotNull())
               .select("key", expr("struct(offset, value) r"))
               .groupBy("key").agg(expr("max(r) r")) 
               .select("r.value"))

    df_read = spark.read.json(
    df_json.rdd.map(lambda x: x.value), multiLine=True)**

Tried with SCALA:

**val df_read = spark.read.json(df_json.rdd.map(x=>x))**

But Im getting below error.

cannot be applied to (org.apache.spark.rdd.RDD[org.apache.spark.sql.Row]) val df_read = spark.read.json(df_json.rdd.map(x=>x))

Any fix? Kindly help.

thebluephantom thebluephantom · Accepted Answer · 2020-08-04T17:07:06

RDD is not supported in Structured Streaming.
Structured Streaming does not allow schema inference.
Schema needs to be defined.

e.g. for a file source

val dataSchema = "Recorded_At timestamp, Device string, Index long, Model string, User string, _corrupt_record String, gt string, x double, y double, z double"
val dataPath = "dbfs:/mnt/training/definitive-guide/data/activity-data-stream.json"

val initialDF = spark
  .readStream                             // Returns DataStreamReader
  .option("maxFilesPerTrigger", 1)        // Force processing of only 1 file per trigger 
  .schema(dataSchema)                     // Required for all streaming DataFrames
  .json(dataPath)                         // The stream's source directory and file type

e.g. Kafka situation as Databricks teach you

spark.conf.set("spark.sql.shuffle.partitions", sc.defaultParallelism)

val kafkaServer = "server1.databricks.training:9092"  // US (Oregon)
// kafkaServer = "server2.databricks.training:9092"   // Singapore

val editsDF = spark.readStream                        // Get the DataStreamReader
  .format("kafka")                                    // Specify the source format as "kafka"
  .option("kafka.bootstrap.servers", kafkaServer)     // Configure the Kafka server name and port
  .option("subscribe", "en")                          // Subscribe to the "en" Kafka topic 
  .option("startingOffsets", "earliest")              // Rewind stream to beginning when we restart notebook
  .option("maxOffsetsPerTrigger", 1000)               // Throttle Kafka's processing of the streams
  .load()                                             // Load the DataFrame
  .select($"value".cast("STRING"))                    // Cast the "value" column to STRING

import org.apache.spark.sql.types.{StructType, StructField, StringType, IntegerType, DoubleType, BooleanType, TimestampType}

lazy val schema = StructType(List(
  StructField("channel", StringType, true),
  StructField("comment", StringType, true),
  StructField("delta", IntegerType, true),
  StructField("flag", StringType, true),
  StructField("geocoding", StructType(List(            //  (OBJECT): Added by the server, field contains IP address geocoding information for anonymous edit.
    StructField("city", StringType, true),
    StructField("country", StringType, true),
    StructField("countryCode2", StringType, true),
    StructField("countryCode3", StringType, true),
    StructField("stateProvince", StringType, true),
    StructField("latitude", DoubleType, true),
    StructField("longitude", DoubleType, true)
  )), true),
  StructField("isAnonymous", BooleanType, true),
  StructField("isNewPage", BooleanType, true),
  StructField("isRobot", BooleanType, true),
  StructField("isUnpatrolled", BooleanType, true),
  StructField("namespace", StringType, true),           //   (STRING): Page's namespace. See https://en.wikipedia.org/wiki/Wikipedia:Namespace 
  StructField("page", StringType, true),                //   (STRING): Printable name of the page that was edited
  StructField("pageURL", StringType, true),             //   (STRING): URL of the page that was edited
  StructField("timestamp", TimestampType, true),        //   (STRING): Time the edit occurred, in ISO-8601 format
  StructField("url", StringType, true),
  StructField("user", StringType, true),                //   (STRING): User who made the edit or the IP address associated with the anonymous editor
  StructField("userURL", StringType, true),
  StructField("wikipediaURL", StringType, true),
  StructField("wikipedia", StringType, true)            //   (STRING): Short name of the Wikipedia that was edited (e.g., "en" for the English)
))

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

val jsonEdits = editsDF.select(
  from_json($"value", schema).as("json")) 
...
...

Spark structured streaming: Schema Inference in Scala

1 Answers