[SPARK-5075] Memory Leak when repartitioning SchemaRDD or running queries in general

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Won't Fix
• Affects Version/s: 1.2.0
• Component/s: PySpark, Spark Core
• Labels: ec2 json memory-leak memory_leak parquet pyspark repartition s3

Environment

spark-ec2 launched 10 node cluster of type c3.8xlarge

• Assignee: Unassigned
• Reporter: Brad Willard
• Created: 04/Jan/15 21:05
• Updated: 22/Apr/15 18:31
• Resolved: 22/Apr/15 18:31

Description

I'm trying to repartition a json dataset for better cpu optimization and save in parquet format for better performance. The Json dataset is about 200gb

from pyspark.sql import SQLContext
sql_context = SQLContext(sc)

rdd = sql_context.jsonFile('s3c://some_path')
rdd = rdd.repartition(256)
rdd.saveAsParquetFile('hdfs://some_path')

In ganglia when the dataset first loads it's about 200G in memory which is expected. However once it attempts the repartition, it balloons over 2.5x in memory which is never returned making any subsequent operations fail from memory errors.

https://s3.amazonaws.com/f.cl.ly/items/3k2n2n3j35273i2v1Y3t/Screen%20Shot%202015-01-04%20at%201.20.29%20PM.png

I'm also seeing a similar memory leak behavior when running repeated queries on a dataset.

rdd = sql_context.parquetFile('hdfs://some_path')
rdd.registerTempTable('events')

sql_context.sql( anything )
sql_context.sql( anything )
sql_context.sql( anything )
sql_context.sql( anything )

will result in a memory usage pattern of.
http://cl.ly/image/180y2D3d1A0X

It seems like intermediate results are not being garbage collected or something. Eventually I have to kill my session to keep running queries.