I want to read nested data from Kafka topics source using Spark Structured Streaming. My Scala code (case classes and Spark processing code):
case class Nested(attr_int: Integer, attr_string: String, attr_float: Float, attr_timestamp: java.sql.Timestamp)
case class Parent(a_str: String, a_long: Long, a_nested: Array[Nested])
import org.apache.spark.sql.Encoders
val jsonSchema = Encoders.product[Parent].schema
val df = sparkSession
.readStream
.format("kafka")
.option("kafka.bootstrap.servers", "localhost:9092")
.option("subscribe", "testnested")
.option("group.id", "testnested")
.option("key.deserializer", "org.apache.kafka.common.serialization.StringSerializer")
.option("value.deserializer", "org.apache.kafka.common.serialization.StringSerializer")
.load()
.select($"value" cast "string" as "json")
.select(from_json($"json", jsonSchema) as "data")
.select("data.*")
.withColumn("nested", explode($"a_nested"))
.select("nested.*")
.as[Nested]
.writeStream
.format("console")
.start()
.awaitTermination()
When I send to Kafka data:
{"a_str":"Str","a_long":100,"a_nested":[{"attr_int":0,"attr_string":"nested_0","attr_float":0.0,"attr_timestamp":"2018-01-01T11:00:00.123321+02:00"},{"attr_int":1,"attr_string":"nested_1","attr_float":1.0,"attr_timestamp":"2018-02-02T12:01:01.023321+02:00"}]}
I get results:
+--------+-----------+----------+--------------------+
|attr_int|attr_string|attr_float| attr_timestamp|
+--------+-----------+----------+--------------------+
| 0| nested_0| 0.0|2018-01-01 13:02:...|
| 1| nested_1| 1.0|2018-02-02 14:01:...|
+--------+-----------+----------+--------------------+
Now I want to get each nested item joined to parent data, f.e.:
+--------+-----------+----------+--------------------+-------+--------+
|attr_int|attr_string|attr_float| attr_timestamp| a_str | a_long |
+--------+-----------+----------+--------------------+-------+--------+
| 0| nested_0| 0.0|2018-01-01 13:02:...| Str | 100 |
| 1| nested_1| 1.0|2018-02-02 14:01:...| Str | 100 |
+--------+-----------+----------+--------------------+-------+--------+
Note that "a_str"
and "a_long"
are columns from the parent entity "Parent"
.
Since I'm not an expert in Spark Structured Streams processing, I want to know what is the most "idiomatic" approach to do it?
Currently I have assumptions:
- Create custom Kafka value deserializer
- Write some kind of join on structured streams (I stuck on it), but I suppose this will require changing json structure (f.e. specify in nested some key value pointing to a parent data)
- Write custom method which will return denormalized data for joined entities and use
flatMap
with this method
Please advise.
Thanks
UPDATE 1: For your convenience I've created correspondent project on GitHub: https://github.com/lospejos/spark-nested-classes-from-json