Does drop_duplicate guarantee to keep the first row and drop rest of the rows after sorting the dataframe in spark?

Question

I have a dataframe, read from Avro file in Hadoop, with three columns (a,b,c), Where one is a key column and among two other columns one is of integer type and the other is of date type.

I am ordering the frame by the integer column and date column and then calling drop_duplicates by key column (a) on the resulting frame.

frame = frame.orderBy(["b","c"],ascending=False)
frame = frame.drop_duplicate('a')

Based on Spark Scala code I can see that orderBy calls sort method internally, which does a global sorting.

/**
   * Returns a new Dataset sorted by the given expressions. For example:
   * {{{
   *   ds.sort($"col1", $"col2".desc)
   * }}}
   *
   * @group typedrel
   * @since 2.0.0
   */
  @scala.annotation.varargs
  def sort(sortExprs: Column*): Dataset[T] = {
    sortInternal(global = true, sortExprs)
  }

https://github.com/apache/spark/blob/branch-2.4/sql/core/src/main/scala/org/apache/spark/sql/Dataset.scala

And Also drop_duplicates(cols) method is translated to Aggregate(first(cols)) as per bellow spark code.

object ReplaceDeduplicateWithAggregate extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transformUpWithNewOutput {
    case d @ Deduplicate(keys, child) if !child.isStreaming =>
      val keyExprIds = keys.map(_.exprId)
      val aggCols = child.output.map { attr =>
        if (keyExprIds.contains(attr.exprId)) {
          attr
        } else {
          Alias(new First(attr).toAggregateExpression(), attr.name)()
        }
      }
      // SPARK-22951: Physical aggregate operators distinguishes global aggregation and grouping
      // aggregations by checking the number of grouping keys. The key difference here is that a
      // global aggregation always returns at least one row even if there are no input rows. Here
      // we append a literal when the grouping key list is empty so that the result aggregate
      // operator is properly treated as a grouping aggregation.
      val nonemptyKeys = if (keys.isEmpty) Literal(1) :: Nil else keys
      val newAgg = Aggregate(nonemptyKeys, aggCols, child)
      val attrMapping = d.output.zip(newAgg.output)
      newAgg -> attrMapping
  }
}

https://github.com/apache/spark/blob/branch-2.4/sql/core/src/main/scala/org/apache/spark/sql/Dataset.scala

So I am expecting that drop duplicate will retrain the first rows after sorting and drop others. but I am observing in my spark jobs that this is not true.

Any Thoughts why?

Answer 1

No.

Sorting by b & c and then dropping by a, would work as you would like, if and only if there was only 1 partition to process. With Big Data that is generally not the case.

So, as you can google elsewhere: dropDuplicates retains the first occurrence of a sort operation - only if there is 1 partition, and otherwise it is luck.

I.e. non-deterministic for when more partitions in play.

Has nothing to do with avro or pyspark. Also, order by b, c may also be non-deterministic.