Coursera Scala Course's Capstone Uses Your Library, but it may not work in that condition.

[codeaperature]

Hi Vincenzo,

To me, it's unclear how to use your library and it's possible that Coursera Scala Course's Capstone (in the build file) has pointed to information that's not longer valid in the readme. I posted this to stackoverflow. This course is hard without being able to do the simple things - it would be nice if you updated your README markdown to help work this issue of TypeTags out. You can note that I tried to make the code on the stackoverflow match Spark's advice, but I also tried to follow the markdown, but didn't post that. In the coursera project, I don't think we can change the build file.

Stefan

[vincenzobaz]

Hi @codeaperature thank you for opening the issue!

To use our encoders, all you need is import scala3encoders.given, then they are available in the implicit scope and you can obtain a reference with summon.

I can adapt your stackoverflow snippets as follows:

import scala3encoders.given
import org.apache.spark.sql.Encoder

case class StationX(stnId: Int, wbanId: Int, lat: Double, lon: Double)

object Station extends App:
  val ss = summon[Encoder[StationX]]
  println(ss.schema)

and

package observatory
import org.apache.spark.sql.SparkSession
import org.apache.spark.sql.Encoder
import scala.reflect.ClassTag
import scala.deriving.Mirror
import scala3udf.{Udf => udf}
import scala3encoders.given

case class CC(i: Int)
object SparkInstance extends App {
  val spark = SparkSession
    .builder()
    .appName("Spark SQL UDF scalar example")
    .getOrCreate()

  def getSchema[T: Mirror.ProductOf: ClassTag] = summon[Encoder[T]].schema
  val random = udf(() => Math.random())
  val plusOne = udf((x: Int) => x + 1)
  val ss = getSchema[CC]
}

You should not need to write a function such as getSchema

[michael72]

I'm a little flustered and worried that an actual course uses spark together with Scala 3 - I would consider this combination experimental and not suited for beginners (although Scala 3 IMHO is much better than Scala 2).

[vincenzobaz]

@michael72 IIRC the course is offered in both Scala 2 and Scala 3.
The assignments were tested in Scala 3 and many students have completed it successfully.

But it has been out for a while, maybe the course manager should investigate whether the scala 3 version has caused more problems...

[codeaperature]

I finally got back to this (I have a regular Data Eng job too) ... I do not believe the parameters of the project mean I can add in extra libraries and it seems that this part does not work in the project:

.../observatory/src/main/scala/observatory/SparkInstance.scala:8:8
Not found: scala3udf
import scala3udf.{Udf => udf}

Maybe I made some other changes. BTW - Did you download the project or just check this in another way?

Since there is no requirement to use Spark and the assignment actually uses a jarred resource ... and per the course suggestion: the data needs to be stream-loaded into memory and then pushed into a spark dataframe/dataset to be processed. I think it's just unnecessary overhead in terms of memory, code and socket open/close time,... I can simply use parallel collections to do a simple join.

I'm going to drop this issue as I am taking a different path, but I am still curious if Coursera provided a bunk suggestion to use your library without supplying the proper tooling in the build.sbt.

Thanks for your past attention to look into this item.

[vincenzobaz]

I think I understand better the issue now. The assignment does not involve udfs, @michael72 implemented the udf a long time after the release of the course.
I could reach out the new person in charge of the courses and tell them to include the udf dependency.

I will also ask if other people reported this issue. I am sorry for the frustration this has caused you.
I collaborated with the course authors so I know it is not easy to maintain a large codebase and still make it extensible.

[codeaperature]

Yeah - I tried to do some things differently ... for example a UDF to convert deg C -> F, but this could be done in another way. Also, I wanted to use datasets with StructTypes automatically derived from case classes.

Thanks for looking into this item for me.