How to display (or operate on) objects encoded by Kryo in Spark Dataset?

北战南征 提交于 2021-02-16 13:54:10

问题


Say you have this:

// assume we handle custom type
class MyObj(val i: Int, val j: String)
implicit val myObjEncoder = org.apache.spark.sql.Encoders.kryo[MyObj]
val ds = spark.createDataset(Seq(new MyObj(1, "a"),new MyObj(2, "b"),new MyObj(3, "c")))

When do a ds.show, I got:

+--------------------+
|               value|
+--------------------+
|[01 00 24 6C 69 6...|
|[01 00 24 6C 69 6...|
|[01 00 24 6C 69 6...|
+--------------------+

I understand that it's because the contents are encoded into internal Spark SQL binary representation. But how can I display the decoded content like this?

+---+---+
| _1| _2|
+---+---+
|  1|  a|
|  2|  b|
|  3|  c|
+---+---+

UPDATE1

Displaying content is not the biggest issue, what's more important is that it could lead to problem when processing the dataset, consider this example:

// continue with the above code
val ds2 = spark.createDataset(Seq(new MyObj(2, "a"),new MyObj(6, "b"),new MyObj(5, "c"))) 

ds.joinWith(ds2, ds("i") === ds2("i"), "inner") 
// this gives a Runtime error: org.apache.spark.sql.AnalysisException: Cannot resolve column name "i" among (value); 

Does this mean, kryo-encoded type is not able to do operation like joinWith conveniently? And how can we process custom type on Dataset then?


回答1:


The following worked for me, but seems like using high-level API to do low-level (deserialization) work.

This is not to say it should be done this way, but shows that it's possible.

I don't know why KryoDeserializer does not deserialize bytes to the object the bytes came from. It is just this way.

One major difference between your class definition and mine is this case that let me using the following trick. Again, no idea exactly why it makes it possible.

scala> println(spark.version)
3.0.1

// Note that case keyword
case class MyObj(val i: Int, val j: String)
import org.apache.spark.sql.Encoders
implicit val myObjEncoder = Encoders.kryo[MyObj]
// myObjEncoder: org.apache.spark.sql.Encoder[MyObj] = class[value[0]: binary]

val ds = (Seq(new MyObj(1, "a"),new MyObj(2, "b"),new MyObj(3, "c"))).toDS
// the Kryo deserializer gives bytes
scala> ds.printSchema
root
 |-- value: binary (nullable = true)

scala> :type sc
org.apache.spark.SparkContext

// Let's deserialize the bytes into an object
import org.apache.spark.serializer.KryoSerializer
val ks = new KryoSerializer(sc.getConf)
// that begs for a generic UDF
val deserMyObj = udf { value: Array[Byte] => 
  import java.nio.ByteBuffer
  ks.newInstance.deserialize(ByteBuffer.wrap(value)).asInstanceOf[MyObj] }

val solution = ds.select(deserMyObj('value) as "result").select($"result.*")
scala> solution.show
+---+---+
|  i|  j|
+---+---+
|  1|  a|
|  2|  b|
|  3|  c|
+---+---+


来源:https://stackoverflow.com/questions/64184387/how-to-display-or-operate-on-objects-encoded-by-kryo-in-spark-dataset

易学教程内所有资源均来自网络或用户发布的内容,如有违反法律规定的内容欢迎反馈
该文章没有解决你所遇到的问题?点击提问,说说你的问题,让更多的人一起探讨吧!