Hi, Amit,

Maybe you can change this configuration spark.sql.shuffle.partitions.
The default is 200 change this property could change the task number when you are using DataFrame API.

On 26 Sep 2017, at 1:25 AM, Amit Sela <amit.sela@venmo.com> wrote:

I'm trying to run a simple pyspark application that reads from file (json), flattens it (explode) and writes back to file (json) partitioned by date using DataFrameWriter.partitionBy(*cols).

I keep getting OOMEs like:
java.lang.OutOfMemoryError: Java heap space
at org.apache.spark.util.collection.unsafe.sort.UnsafeSorterSpillWriter.<init>(UnsafeSorterSpillWriter.java:46)
at org.apache.spark.util.collection.unsafe.sort.UnsafeExternalSorter.spill(UnsafeExternalSorter.java:206)
at org.apache.spark.memory.TaskMemoryManager.acquireExecutionMemory(TaskMemoryManager.java:203)

Explode could make the underlying RDD grow a lot, and maybe in an unbalanced way sometimes,  
adding to that partitioning by date (in daily ETLs for instance) would probably cause a data skew (right?), but why am I getting OOMs? Isn't Spark supposed to spill to disk if the underlying RDD is too big to fit in memory?

If I'm not using "partitionBy" with the writer (still exploding) everything works fine.

This happens both in EMR and in local (mac) pyspark/spark shell (tried both in python and scala).