suggestion for tune up the code which contains explode and groupby

Question

I wrote the code for below probelem but it has below problems. Please suggest me if some tuning can be done.

It takes more time I think.
there are 3 brands as of now. It is hardcoded. If more brands would be added, i need to add the code manually.

input dataframe schema :

root
 |-- id: string (nullable = true)
 |-- attrib: map (nullable = true)
 |    |-- key: string
 |    |-- value: string (valueContainsNull = true)
 |-- pref: array (nullable = true)
 |    |-- element: struct (containsNull = true)
 |    |    |-- pref_type: string (nullable = true)
 |    |    |-- brand: string (nullable = true)
 |    |    |-- tp_id: string (nullable = true)
 |    |    |-- aff: float (nullable = true)
 |    |    |-- pre_id: string (nullable = true)
 |    |    |-- cr_date: string (nullable = true)
 |    |    |-- up_date: string (nullable = true)
 |    |    |-- pref_attrib: map (nullable = true)
 |    |    |    |-- key: string
 |    |    |    |-- value: string (valueContainsNull = true)

expected output schema:

root
 |-- id: string (nullable = true)
 |-- attrib: map (nullable = true)
 |    |-- key: string
 |    |-- value: string (valueContainsNull = true)
 |-- pref: struct (nullable = false)
 |    |-- brandA: array (nullable = true)
 |    |    |-- element: struct (containsNull = false)
 |    |    |    |-- pref_type: string (nullable = true)
 |    |    |    |-- tp_id: string (nullable = true)
 |    |    |    |-- aff: float (nullable = true)
 |    |    |    |-- pref_id: string (nullable = true)
 |    |    |    |-- cr_date: string (nullable = true)
 |    |    |    |-- up_date: string (nullable = true)
 |    |    |    |-- pref_attrib: map (nullable = true)
 |    |    |    |    |-- key: string
 |    |    |    |    |-- value: string (valueContainsNull = true)
 |    |-- brandB: array (nullable = true)
 |    |    |-- element: struct (containsNull = false)
 |    |    |    |-- pref_type: string (nullable = true)
 |    |    |    |-- tp_id: string (nullable = true)
 |    |    |    |-- aff: float (nullable = true)
 |    |    |    |-- pref_id: string (nullable = true)
 |    |    |    |-- cr_date: string (nullable = true)
 |    |    |    |-- up_date: string (nullable = true)
 |    |    |    |-- pref_attrib: map (nullable = true)
 |    |    |    |    |-- key: string
 |    |    |    |    |-- value: string (valueContainsNull = true)
 |    |-- brandC: array (nullable = true)
 |    |    |-- element: struct (containsNull = false)
 |    |    |    |-- pref_type: string (nullable = true)
 |    |    |    |-- tp_id: string (nullable = true)
 |    |    |    |-- aff: float (nullable = true)
 |    |    |    |-- pref_id: string (nullable = true)
 |    |    |    |-- cr_date: string (nullable = true)
 |    |    |    |-- up_date: string (nullable = true)
 |    |    |    |-- pref_attrib: map (nullable = true)
 |    |    |    |    |-- key: string
 |    |    |    |    |-- value: string (valueContainsNull = true)

The processing can be done based on the brand attribute under preferences(preferences.brand)

I have written the below code for that:

def modifyBrands(inputDf: DataFrame): DataFrame ={
    val PreferenceProps = Array("pref_type", "tp_id", "aff", "pref_id", "cr_date", "up_date", "pref_attrib")
    import org.apache.spark.sql.functions._
    val explodedDf = inputDf.select(col("id"), explode(col("pref")))
        .select(
            col("id"),
            col("col.pref_type"),
            col("col.brand"),
            col("col.tp_id"),
            col("col.aff"),
            col("col.pre_id"),
            col("col.cr_dt"),
            col("col.up_dt"),
            col("col.pref_attrib")
        ).cache()

    val brandAddedDf = explodedDf
        .withColumn("brandA", when(col("brand") === "brandA", struct(PreferenceProps.head, PreferenceProps.tail:_*)).as("brandA"))
        .withColumn("brandB", when(col("brand") === "brandB", struct(PreferenceProps.head, PreferenceProps.tail:_*)).as("brandB"))
        .withColumn("brandC", when(col("brand") === "brandC", struct(PreferenceProps.head, PreferenceProps.tail:_*)).as("brandC"))
        .cache()

    explodedDf.unpersist()

    val groupedDf = brandAddedDf.groupBy("id").agg(
        collect_list("brandA").alias("brandA"),
        collect_list("brandB").alias("brandB"),
        collect_list("brandC").alias("brandC")
    ).withColumn("preferences", struct(
        when(size(col("brandA")).notEqual(0), col("brandA")).alias("brandA"),
        when(size(col("brandB")).notEqual(0), col("brandB")).alias("brandB"),
        when(size(col("brandC")).notEqual(0), col("brandC")).alias("brandC"),
    )).drop("brandA", "brandB", "brandC")
      .cache()
    brandAddedDf.unpersist()

    val idAttributesDf = inputDf.select("id", "attrib").cache()

    val joinedDf = idAttributesDf.join(groupedDf, "id")
    groupedDf.unpersist()
    idAttributesDf.unpersist()
    joinedDf.printSchema()
    joinedDf // returning joined df which will be wrote as paquet file.
  }

user_1386551 · Answer

You can simplify your code using higher-order function filter on arrays. Just map through brand names and for-each one return a filtered array from pref. This way you avoid the exploding / grouping part.

Here's a complete example:

val data = """{"id":1,"attrib":{"key":"k","value":"v"},"pref":[{"pref_type":"type1","brand":"brandA","tp_id":"id1","aff":"aff1","pre_id":"pre_id1","cr_date":"2021-01-06","up_date":"2021-01-06","pref_attrib":{"key":"k","value":"v"}},{"pref_type":"type1","brand":"brandB","tp_id":"id1","aff":"aff1","pre_id":"pre_id1","cr_date":"2021-01-06","up_date":"2021-01-06","pref_attrib":{"key":"k","value":"v"}},{"pref_type":"type1","brand":"brandC","tp_id":"id1","aff":"aff1","pre_id":"pre_id1","cr_date":"2021-01-06","up_date":"2021-01-06","pref_attrib":{"key":"k","value":"v"}}]}"""
val inputDf = spark.read.json(Seq(data).toDS)

val brands = Seq("brandA", "brandB", "brandC")

// or getting them from input dataframe
// val brands = inputDf.select("pref.brand").as[Seq[String]].collect.flatten


val brandAddedDf = inputDf.withColumn(
  "pref",
  struct(brands.map(b => expr(s"filter(pref, x -> x.brand = '$b')").as(b)): _*)
)

brandAddedDf.printSchema
//root
// |-- attrib: struct (nullable = true)
// |    |-- key: string (nullable = true)
// |    |-- value: string (nullable = true)
// |-- id: long (nullable = true)
// |-- pref: struct (nullable = false)
// |    |-- brandA: array (nullable = true)
// |    |    |-- element: struct (containsNull = true)
// |    |    |    |-- aff: string (nullable = true)
// |    |    |    |-- brand: string (nullable = true)
// |    |    |    |-- cr_date: string (nullable = true)
// |    |    |    |-- pre_id: string (nullable = true)
// |    |    |    |-- pref_attrib: struct (nullable = true)
// |    |    |    |    |-- key: string (nullable = true)
// |    |    |    |    |-- value: string (nullable = true)
// |    |    |    |-- pref_type: string (nullable = true)
// |    |    |    |-- tp_id: string (nullable = true)
// |    |    |    |-- up_date: string (nullable = true)
// |    |-- brandB: array (nullable = true)
// |    |    |-- element: struct (containsNull = true)
// |    |    |    |-- aff: string (nullable = true)
// |    |    |    |-- brand: string (nullable = true)
// |    |    |    |-- cr_date: string (nullable = true)
// |    |    |    |-- pre_id: string (nullable = true)
// |    |    |    |-- pref_attrib: struct (nullable = true)
// |    |    |    |    |-- key: string (nullable = true)
// |    |    |    |    |-- value: string (nullable = true)
// |    |    |    |-- pref_type: string (nullable = true)
// |    |    |    |-- tp_id: string (nullable = true)
// |    |    |    |-- up_date: string (nullable = true)
// |    |-- brandC: array (nullable = true)
// |    |    |-- element: struct (containsNull = true)
// |    |    |    |-- aff: string (nullable = true)
// |    |    |    |-- brand: string (nullable = true)
// |    |    |    |-- cr_date: string (nullable = true)
// |    |    |    |-- pre_id: string (nullable = true)
// |    |    |    |-- pref_attrib: struct (nullable = true)
// |    |    |    |    |-- key: string (nullable = true)
// |    |    |    |    |-- value: string (nullable = true)
// |    |    |    |-- pref_type: string (nullable = true)
// |    |    |    |-- tp_id: string (nullable = true)
// |    |    |    |-- up_date: string (nullable = true)

user_13535120 · Answer

I think they're are a couple issues with how you are doing your code, but the real way to tell where you have a problem with your code is to look at the SPARK UI. I find the "Jobs" tab and the "SQL" tab very informative to figure out where the code is spending most of its time. Then see if those parts can be re-written to give you more speed. Some of the items I point out below may not matter if there is a bottleneck elsewhere that really is where most of the time is being spent.

There are reasons to create nested structures (Like you are for Brand). I'm just not sure I see the payoff here and it's not explained. It should be considered why you are maintaining this structure and what the benefit is. Is there a performance gain for maintaining it? Or is it simply an artifact of how the data was created?

General tips that might help a little:

In general you should only cache code that you will use more than once. You have a lot of code you don't use more than once but you still cache.

Small, small performance boost. (So in other words when you need every millisecond...) withColumn actually doesn't perform as well as select. (Likely due to some object creation) where possible use select instead of withColumn. Not really worth re-writing your code unless you really need every milli-second.

suggestion for tune up the code which contains explode and groupby

2 Answers