Partitioning a large skewed dataset in S3 with Spark's partitionBy method

Question

I am trying to write out a large partitioned dataset to disk with Spark and the partitionBy algorithm is struggling with both of the approaches I\'ve tried.

Answer 1

The simplest solution is to add one or more columns to repartition and explicitly set the number of partitions.

val numPartitions = ???

df.repartition(numPartitions, $"some_col", $"some_other_col")
 .write.partitionBy("some_col")
 .parquet("partitioned_lake")

where:

• numPartitions - should be an upper bound (actual number can be lower) of the desired number of files written to a partition directory.

• $"some_other_col" (and optional additional columns) should have high cardinality and be independent of the $"some_column (there should be functional dependency between these two, and shouldn't be highly correlated).

  If data doesn't contain such column you can use o.a.s.sql.functions.rand.

  import org.apache.spark.sql.functions.rand
  
  df.repartition(numPartitions, $"some_col", rand)
    .write.partitionBy("some_col")
    .parquet("partitioned_lake")