Scala Spark: Flatten Array of Key/Value structs

I wouldn't worry too much about grouping by several columns, other than potentially making things confusing. In that vein, if there is a simpler, more maintainable way, go for it. Without example input/output, I'm not sure if this gets you where you're trying to go, but maybe it'll be of use:

Seq(Seq("k1" -> "v1", "k2" -> "v2")).toDS() // some basic input based on my understanding of your description .select(explode($"value")) // flatten the array .select("col.*") // de-nest the struct .groupBy("_2") // one row per distinct value .pivot("_1") // one column per distinct key .count // or agg(first) if you want the value in each column .show +---+----+----+ | _2| k1| k2| +---+----+----+ | v2|null| 1| | v1| 1|null| +---+----+----+ 

Based on what you've now said, I get the impression that there are many columns like "state" that aren't required for the aggregation, but need to be in the final result.

For reference, if you didn't need to pivot, you could add a struct column with all such fields nested within, then add it to your aggregation, eg: .agg(first($"myStruct"), first($"number")). The main advantage is only having actual key column(s) referenced in the groubBy. But when using pivot things get a little weird, so we'll set that option aside.

In this use case, the simplest way I could come up with involves splitting your dataframe and joining it back together after the aggregation using some rowkey. In this example I am assuming that "index" is suitable for that purpose:

 val mehCols = dfExampleInput.columns.filter(_ != "entries").map(col) val mehDF = dfExampleInput.select(mehCols:_*) val aggDF = dfExampleInput .select($"index", explode($"entries").as("entry")) .select($"index", $"entry.*") .groupBy("index") .pivot("name") .agg(first($"number")) scala> mehDF.join(aggDF, Seq("index")).show +-----+-----+---+----+ |index|state| A| B| +-----+-----+---+----+ | 0| KY| 45|null| | 1| OR| 30| 10| +-----+-----+---+----+ 

I doubt you would see much of a difference in performance, if any. Maybe at the extremes, eg: very many meh columns, or very many pivot columns, or something like that, or maybe nothing at all. Personally, I would test both with decently-sized input, and if there wasn't a significant difference, use whichever one seemed easier to maintain.

Here is another way that is based on the assumption that there are no duplicates on the entries column i.e Seq(testStruct("A", "30"), testStruct("A", "70"), testStruct("B", "10")) will cause an error. The next solution combines both RDD and Dataframe APIs for the implementation:

import org.apache.spark.sql.functions.explode import org.apache.spark.sql.types.StructType case class testStruct(name : String, number : String) val df = Seq( (0, "KY", Seq(testStruct("A", "45"))), (1, "OR", Seq(testStruct("A", "30"), testStruct("B", "10"))), (2, "FL", Seq(testStruct("A", "30"), testStruct("B", "10"), testStruct("C", "20"))), (3, "TX", Seq(testStruct("B", "60"), testStruct("A", "19"), testStruct("C", "40"))) ) .toDF("index", "state", "entries") .cache // get all possible keys from entries i.e Seq[A, B, C] val finalCols = df.select(explode($"entries").as("entry")) .select($"entry".getField("name").as("entry_name")) .distinct .collect .map{_.getAs[String]("entry_name")} .sorted // Attention: we need to retain the order of the columns // 1. when generating row values and // 2. when creating the schema val rdd = df.rdd.map{ r => // transform the entries array into a map i.e Map(A -> 30, B -> 10) val entriesMap = r.getSeq[Row](2).map{r => (r.getString(0), r.getString(1))}.toMap // transform finalCols into a map with null value i.e Map(A -> null, B -> null, C -> null) val finalColsMap = finalCols.map{c => (c, null)}.toMap // replace null values with those that are present from the current row by merging the two previous maps // Attention: this should retain the order of finalColsMap val merged = finalColsMap ++ entriesMap // concatenate the two first row values ["index", "state"] with the values from merged val finalValues = Seq(r(0), r(1)) ++ merged.values Row.fromSeq(finalValues) } val extraCols = finalCols.map{c => s"`${c}` STRING"} val schema = StructType.fromDDL("`index` INT, `state` STRING," + extraCols.mkString(",")) val finalDf = spark.createDataFrame(rdd, schema) finalDf.show // +-----+-----+---+----+----+ // |index|state| A| B| C| // +-----+-----+---+----+----+ // | 0| KY| 45|null|null| // | 1| OR| 30| 10|null| // | 2| FL| 30| 10| 20| // | 3| TX| 19| 60| 40| // +-----+-----+---+----+----+ 

Note: the solution requires one extra action to retrieve the unique keys although it doesn't cause any shuffling since it it based on narrow transformations only.

I've worked out a solution myself:

def extractFromArray(colName : String, key : String, numKeys : Int, keyName : String) = { val indexCols = (0 to numKeys-1).map(col(colName).getItem(_)) indexCols.foldLeft(lit(null))((innerCol : Column, indexCol : Column) => when(indexCol.isNotNull && (indexCol.getItem(keyName) === key), indexCol) .otherwise(innerCol)) } 

Example:

case class testStruct(name : String, number : String) val df = Seq( (0, "KY", Seq(testStruct("A", "45"))), (1, "OR", Seq(testStruct("A", "30"), testStruct("B", "10"))), (2, "FL", Seq(testStruct("A", "30"), testStruct("B", "10"), testStruct("C", "20"))), (3, "TX", Seq(testStruct("B", "60"), testStruct("A", "19"), testStruct("C", "40"))) ) .toDF("index", "state", "entries") .withColumn("A", extractFromArray("entries", "B", 3, "name")) .show 

which produces:

+-----+-----+--------------------+-------+ |index|state| entries| A| +-----+-----+--------------------+-------+ | 0| KY| [[A, 45]]| null| | 1| OR| [[A, 30], [B, 10]]|[B, 10]| | 2| FL|[[A, 30], [B, 10]...|[B, 10]| | 3| TX|[[B, 60], [A, 19]...|[B, 60]| +-----+-----+--------------------+-------+ 

This solution is a little different from other answers:

• It works on only a single key at a time
• It requires the key name and number of keys be known in advance
• It produces a column of structs, rather than doing the extra step of extracting specific values
• It works as a simple column-to-column operation, rather than requiring transformations on the entire DF
• It can be evaluated lazily

The first three issues can be handled by calling code, and leave it somewhat more flexible for cases where you already know the keys or where the structs contain additional values to extract.