[SPARK-23246] (Py)Spark OOM because of iteratively accumulated metadata that cannot be cleared

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Not A Problem
• Affects Version/s: 2.2.1
• Component/s: PySpark, Spark Core, SQL
• Assignee: Unassigned
• Reporter: V Luong
• Created: 27/Jan/18 04:50
• Updated: 30/Jan/18 03:43
• Resolved: 29/Jan/18 23:56

Description

I am having consistent OOM crashes when trying to use PySpark for iterative algorithms in which I create new DataFrames per iteration (e.g. by sampling from a "mother" DataFrame), do something with such DataFrames, and never need such DataFrames ever in future iterations.

The below script simulates such OOM failures. Even when one tries explicitly .unpersist() the temporary DataFrames (by using the --unpersist flag below) and/or deleting and garbage-collecting the Python objects (by using the --py-gc flag below), the Java objects seem to stay on and accumulate until they exceed the JVM/driver memory.

The more complex the temporary DataFrames in each iteration (illustrated by the --n-partitions flag below), the faster OOM occurs.

The typical error messages include:

• "java.lang.OutOfMemoryError : GC overhead limit exceeded"

• "Java heap space"
• "ERROR TransportRequestHandler: Error sending result RpcResponse{requestId=6053742323219781
  161, body=NioManagedBuffer{buf=java.nio.HeapByteBuffer[pos=0 lim=47 cap=64]}} to /<IP ADDR>; closing connection"

Please suggest how I may overcome this so that we can have long-running iterative programs using Spark that uses resources only up to a bounded, controllable limit.

 

from __future__ import print_function

import argparse
import gc
import pandas

import pyspark


arg_parser = argparse.ArgumentParser()
arg_parser.add_argument('--unpersist', action='store_true')
arg_parser.add_argument('--py-gc', action='store_true')
arg_parser.add_argument('--n-partitions', type=int, default=1000)
args = arg_parser.parse_args()


# create SparkSession (*** set spark.driver.memory to 512m in spark-defaults.conf ***)
spark = pyspark.sql.SparkSession.builder \
    .config('spark.executor.instances', 2) \
    .config('spark.executor.cores', 2) \
    .config('spark.executor.memory', '512m') \
    .config('spark.ui.enabled', False) \
    .config('spark.ui.retainedJobs', 10) \
    .config('spark.ui.retainedStages', 10) \
    .config('spark.ui.retainedTasks', 10) \
    .enableHiveSupport() \
    .getOrCreate()


# create Parquet file for subsequent repeated loading
df = spark.createDataFrame(
    pandas.DataFrame(
        dict(
            row=range(args.n_partitions),
            x=args.n_partitions * [0]
        )
    )
)

parquet_path = '/tmp/TestOOM-{}Partitions.parquet'.format(args.n_partitions)

df.write.parquet(
    path=parquet_path,
    partitionBy='row',
    mode='overwrite'
)


i = 0


# the below loop simulates an iterative algorithm that creates new DataFrames in each iteration (e.g. sampling from a "mother" DataFrame), do something, and never need those DataFrames again in future iterations
# we are having a problem cleaning up the built-up metadata
# hence the program will crash after while because of OOM
while True:
    _df = spark.read.parquet(parquet_path)

    if args.unpersist:
        _df.unpersist()

    if args.py_gc:
        del _df
        gc.collect()

    i += 1; print('COMPLETED READ ITERATION #{}\n'.format(i))