AWS Glue job failing with OOM exception when changing column names - pyspark

Question

I have an ETL job where I load some data from S3 into a dynamic frame, relationalize it, and iterate through the dynamic frames returned. I want to query the result of this in Athena later so I want to change the names of the columns from having '.' to '_' and lower case them. When I do this transformation, I change the DynamicFrame into a spark dataframe and have been doing it this way. I've also seen a problem in another SO question where it turned out there is a reported problem with AWS Glue rename field transform so I've stayed away from that.
I've tried a couple things, including adding a load limit size to 50MB, repartitioning the dataframe, using both dataframe.schema.names and dataframe.columns, using reduce instead of loops, using sparksql to change it and nothing has worked. I'm fairly certain that its this transformation that failing because I've put some print statements in and the print that I have right after the completion of this transformation never shows up. I used a UDF at one point but that also failed. I've tried the actual transformation using df.toDF(new_column_names) and df.withColumnRenamed() but it never gets this far because I've not seen it get past retrieving the column names. Here's the code I've been using. I've been changing the actual name transformation as I said above, but the rest of it has stayed pretty much the same.
I've seen some people try and use the spark.executor.memory, spark.driver.memory, spark.executor.memoryOverhead and spark.driver.memoryOverhead. I've used those and set them to the most AWS Glue will let you but to no avail.
import sys
from awsglue.transforms import *
from awsglue.utils import getResolvedOptions
from pyspark.context import SparkContext
from awsglue.context import GlueContext
from awsglue.job import Job
from awsglue.dynamicframe import DynamicFrame
from pyspark.sql.functions import explode, col, lower, trim, regexp_replace
import copy
import json
import boto3
import botocore
import time
# ========================================================
# UTILITY FUNCTIONS
# ========================================================
def lower_and_pythonize(s=None):
if s is not None:
return s.replace('.', '_').lower()
else:
return None
# pyspark implementation of renaming
# exprs = [
# regexp_replace(lower(trim(col(c))),'\.' , '_').alias(c) if t == "string" else col(c)
# for (c, t) in data_frame.dtypes
# ]
# ========================================================
# END UTILITY FUNCTIONS
# ========================================================
## #params: [JOB_NAME]
args = getResolvedOptions(sys.argv, ['JOB_NAME'])
sc = SparkContext()
glueContext = GlueContext(sc)
spark = glueContext.spark_session
job = Job(glueContext)
job.init(args['JOB_NAME'], args)
#my params
bucket_name = '<my-s3-bucket>' # name of the bucket. do not include 's3://' thats added later
output_key = '<my-output-path>' # key where all of the output is saved
input_keys = ['<root-directory-i'm using'] # highest level key that holds all of the desired data
s3_exclusions = "[\"*.orc\"]" # list of strings to exclude. Documentation: https://docs.aws.amazon.com/glue/latest/dg/aws-glue-programming-etl-connect.html#aws-glue-programming-etl-connect-s3
s3_exclusions = s3_exclusions.replace('\n', '')
dfc_root_table_name = 'root' # name of the root table generated in the relationalize process
input_paths = ['s3://' + bucket_name + '/' + x for x in input_keys] # turn input keys into s3 paths
output_connection_opts = {"path": "s3://" + bucket_name + "/" + output_key} # dict of options. Documentation link found above the write_dynamic_frame.from_options line
s3_client = boto3.client('s3', 'us-east-1') # s3 client used for writing to s3
s3_resource = boto3.resource('s3', 'us-east-1') # s3 resource used for checking if key exists
group_mb = 50 # NOTE: 75 has proven to be too much when running on all of the april data
group_size = str(group_mb * 1024 * 1024)
input_connection_opts = {'paths': input_paths,
'groupFiles': 'inPartition',
'groupSize': group_size,
'recurse': True,
'exclusions': s3_exclusions} # dict of options. Documentation link found above the create_dynamic_frame_from_options line
print(sc._conf.get('spark.executor.cores'))
num_paritions = int(sc._conf.get('spark.executor.cores')) * 4
print('Loading all json files into DynamicFrame...')
loading_time = time.time()
df = glueContext.create_dynamic_frame_from_options(connection_type='s3', connection_options=input_connection_opts, format='json')
print('Done. Time to complete: {}s'.format(time.time() - loading_time))
# using the list of known null fields (at least on small sample size) remove them
#df = df.drop_fields(drop_paths)
# drop any remaining null fields. The above covers known problems that this step doesn't fix
print('Dropping null fields...')
dropping_time = time.time()
df_without_null = DropNullFields.apply(frame=df, transformation_ctx='df_without_null')
print('Done. Time to complete: {}s'.format(time.time() - dropping_time))
df = None
print('Relationalizing dynamic frame...')
relationalizing_time = time.time()
dfc = Relationalize.apply(frame=df_without_null, name=dfc_root_table_name, info="RELATIONALIZE", transformation_ctx='dfc', stageThreshold=3)
print('Done. Time to complete: {}s'.format(time.time() - relationalizing_time))
keys = dfc.keys()
keys.sort(key=lambda s: len(s))
print('Writting all dynamic frames to s3...')
writting_time = time.time()
for key in keys:
good_key = lower_and_pythonize(s=key)
data_frame = dfc.select(key).toDF()
# lowercase all the names and remove '.'
print('Removing . and _ from names for {} frame...'.format(key))
df_fix_names_time = time.time()
print('Repartitioning data frame...')
data_frame.repartition(num_paritions)
print('Done.')
#
print('Changing names...')
for old_name in data_frame.schema.names:
data_frame = data_frame.withColumnRenamed(old_name, old_name.replace('.','_').lower())
print('Done.')
#
df_now = DynamicFrame.fromDF(dataframe=data_frame, glue_ctx=glueContext, name='df_now')
print('Done. Time to complete: {}'.format(time.time() - df_fix_names_time))
# if a conflict of types appears, make it 2 columns
# https://docs.aws.amazon.com/glue/latest/dg/built-in-transforms.html
print('Fixing any type conficts for {} frame...'.format(key))
df_resolve_time = time.time()
resolved = ResolveChoice.apply(frame = df_now, choice = 'make_cols', transformation_ctx = 'resolved')
print('Done. Time to complete: {}'.format(time.time() - df_resolve_time))
# check if key exists in s3. if not make one
out_connect = copy.deepcopy(output_connection_opts)
out_connect['path'] = out_connect['path'] + '/' + str(good_key)
try:
s3_resource.Object(bucket_name, output_key + '/' + good_key + '/').load()
except botocore.exceptions.ClientError as e:
if e.response['Error']['Code'] == '404' or 'NoSuchKey' in e.response['Error']['Code']:
# object doesn't exist
s3_client.put_object(Bucket=bucket_name, Key=output_key+'/'+good_key + '/')
else:
print(e)
## https://docs.aws.amazon.com/glue/latest/dg/aws-glue-api-crawler-pyspark-extensions-glue-context.html
print('Writing {} frame to S3...'.format(key))
df_writing_time = time.time()
datasink4 = glueContext.write_dynamic_frame.from_options(frame = df_now, connection_type = "s3", connection_options = out_connect, format = "orc", transformation_ctx = "datasink4")
out_connect = None
datasink4 = None
print('Done. Time to complete: {}'.format(time.time() - df_writing_time))
print('Done. Time to complete: {}s'.format(time.time() - writting_time))
job.commit()
Here is the error I'm getting
19/06/07 16:33:36 DEBUG Client:
client token: N/A
diagnostics: Application application_1559921043869_0001 failed 1 times due to AM Container for appattempt_1559921043869_0001_000001 exited with exitCode: -104
For more detailed output, check application tracking page:http://ip-172-32-9-38.ec2.internal:8088/cluster/app/application_1559921043869_0001Then, click on links to logs of each attempt.
Diagnostics: Container [pid=9630,containerID=container_1559921043869_0001_01_000001] is running beyond physical memory limits. Current usage: 5.6 GB of 5.5 GB physical memory used; 8.8 GB of 27.5 GB virtual memory used. Killing container.
Dump of the process-tree for container_1559921043869_0001_01_000001 :
|- PID PPID PGRPID SESSID CMD_NAME USER_MODE_TIME(MILLIS) SYSTEM_TIME(MILLIS) VMEM_USAGE(BYTES) RSSMEM_USAGE(PAGES) FULL_CMD_LINE
|- 9630 9628 9630 9630 (bash) 0 0 115822592 675 /bin/bash -c LD_LIBRARY_PATH=/usr/lib/hadoop/lib/native:/usr/lib/hadoop-lzo/lib/native:::/usr/lib/hadoop-lzo/lib/native:/usr/lib/hadoop/lib/native::/usr/lib/hadoop-lzo/lib/native:/usr/lib/hadoop/lib/native:/usr/lib/hadoop-lzo/lib/native:/usr/lib/hadoop/lib/native /usr/lib/jvm/java-openjdk/bin/java -server -Xmx5120m -Djava.io.tmpdir=/mnt/yarn/usercache/root/appcache/application_1559921043869_0001/container_1559921043869_0001_01_000001/tmp '-XX:+UseConcMarkSweepGC' '-XX:CMSInitiatingOccupancyFraction=70' '-XX:MaxHeapFreeRatio=70' '-XX:+CMSClassUnloadingEnabled' '-XX:OnOutOfMemoryError=kill -9 %p' '-Djavax.net.ssl.trustStore=ExternalAndAWSTrustStore.jks' '-Djavax.net.ssl.trustStoreType=JKS' '-Djavax.net.ssl.trustStorePassword=amazon' '-DRDS_ROOT_CERT_PATH=rds-combined-ca-bundle.pem' '-DREDSHIFT_ROOT_CERT_PATH=redshift-ssl-ca-cert.pem' '-DRDS_TRUSTSTORE_URL=file:RDSTrustStore.jks' -Dspark.yarn.app.container.log.dir=/var/log/hadoop-yarn/containers/application_1559921043869_0001/container_1559921043869_0001_01_000001 org.apache.spark.deploy.yarn.ApplicationMaster --class 'org.apache.spark.deploy.PythonRunner' --primary-py-file runscript.py --arg 'script_2019-06-07-15-29-50.py' --arg '--JOB_NAME' --arg 'tss-json-to-orc' --arg '--JOB_ID' --arg 'j_f9f7363e5d8afa20784bc83d7821493f481a78352641ad2165f8f68b88c8e5fe' --arg '--JOB_RUN_ID' --arg 'jr_a77087792dd74231be1f68c1eda2ed33200126b8952c5b1420cb6684759cf233' --arg '--job-bookmark-option' --arg 'job-bookmark-disable' --arg '--TempDir' --arg 's3://aws-glue-temporary-059866946490-us-east-1/zmcgrath' --properties-file /mnt/yarn/usercache/root/appcache/application_1559921043869_0001/container_1559921043869_0001_01_000001/__spark_conf__/__spark_conf__.properties 1> /var/log/hadoop-yarn/containers/application_1559921043869_0001/container_1559921043869_0001_01_000001/stdout 2> /var/log/hadoop-yarn/containers/application_1559921043869_0001/container_1559921043869_0001_01_000001/stderr
|- 9677 9648 9630 9630 (python) 12352 2628 1418354688 261364 python runscript.py script_2019-06-07-15-29-50.py --JOB_NAME tss-json-to-orc --JOB_ID j_f9f7363e5d8afa20784bc83d7821493f481a78352641ad2165f8f68b88c8e5fe --JOB_RUN_ID jr_a77087792dd74231be1f68c1eda2ed33200126b8952c5b1420cb6684759cf233 --job-bookmark-option job-bookmark-disable --TempDir s3://aws-glue-temporary-059866946490-us-east-1/zmcgrath
|- 9648 9630 9630 9630 (java) 265906 3083 7916974080 1207439 /usr/lib/jvm/java-openjdk/bin/java -server -Xmx5120m -Djava.io.tmpdir=/mnt/yarn/usercache/root/appcache/application_1559921043869_0001/container_1559921043869_0001_01_000001/tmp -XX:+UseConcMarkSweepGC -XX:CMSInitiatingOccupancyFraction=70 -XX:MaxHeapFreeRatio=70 -XX:+CMSClassUnloadingEnabled -XX:OnOutOfMemoryError=kill -9 %p -Djavax.net.ssl.trustStore=ExternalAndAWSTrustStore.jks -Djavax.net.ssl.trustStoreType=JKS -Djavax.net.ssl.trustStorePassword=amazon -DRDS_ROOT_CERT_PATH=rds-combined-ca-bundle.pem -DREDSHIFT_ROOT_CERT_PATH=redshift-ssl-ca-cert.pem -DRDS_TRUSTSTORE_URL=file:RDSTrustStore.jks -Dspark.yarn.app.container.log.dir=/var/log/hadoop-yarn/containers/application_1559921043869_0001/container_1559921043869_0001_01_000001 org.apache.spark.deploy.yarn.ApplicationMaster --class org.apache.spark.deploy.PythonRunner --primary-py-file runscript.py --arg script_2019-06-07-15-29-50.py --arg --JOB_NAME --arg tss-json-to-orc --arg --JOB_ID --arg j_f9f7363e5d8afa20784bc83d7821493f481a78352641ad2165f8f68b88c8e5fe --arg --JOB_RUN_ID --arg jr_a77087792dd74231be1f68c1eda2ed33200126b8952c5b1420cb6684759cf233 --arg --job-bookmark-option --arg job-bookmark-disable --arg --TempDir --arg s3://aws-glue-temporary-059866946490-us-east-1/zmcgrath --properties-file /mnt/yarn/usercache/root/appcache/application_1559921043869_0001/container_1559921043869_0001_01_000001/__spark_conf__/__spark_conf__.properties
Container killed on request. Exit code is 143
Container exited with a non-zero exit code 143
Failing this attempt. Failing the application.
ApplicationMaster host: N/A
ApplicationMaster RPC port: -1
queue: default
start time: 1559921462650
final status: FAILED
tracking URL: http://ip-172-32-9-38.ec2.internal:8088/cluster/app/application_1559921043869_0001
user: root
Here are the log contents from the job
LogType:stdout
Log Upload Time:Fri Jun 07 16:33:36 +0000 2019
LogLength:487
Log Contents:
4
Loading all json files into DynamicFrame...
Done. Time to complete: 59.5056920052s
Dropping null fields...
null_fields [<some fields that were dropped>]
Done. Time to complete: 529.95293808s
Relationalizing dynamic frame...
Done. Time to complete: 2773.11689401s
Writting all dynamic frames to s3...
Removing . and _ from names for root frame...
Repartitioning data frame...
Done.
Changing names...
End of LogType:stdout
As I said earlier, the Done. print after changing the names never appears in the logs. I've seen plenty of people getting the same error I'm seeing and I've tried a fair bit of them with no success. Any help you can provide would b e much appreciated. Let me know if you need any more information. Thanks
Edit
Prabhakar's comment reminded me that I have tried the memory worker type in AWS Glue and it still failed. As stated above, I have tried raising the amount of memory in the memoryOverhead from 5 to 12, but to avail. Neither of these made the job complete successfully
Update
I put in the following code for column name change instead of the above code for easier debugging
print('Changing names...')
name_counter = 0
for old_name in data_frame.schema.names:
print('Name number {}. name being changed: {}'.format(name_counter, old_name))
data_frame = data_frame.withColumnRenamed(old_name, old_name.replace('.','_').lower())
name_counter += 1
print('Done.')
And I got the following output
Removing . and _ from names for root frame...
Repartitioning data frame...
Done.
Changing names...
End of LogType:stdout
So it must be a problem with the data_frame.schema.names part. Could it be this line with my loop through all of the DynamicFrames? Am I looping through the DynamicFrames from the relationalize transformation correctly?
Update 2
Glue recently added more verbose logs and I found this
ERROR YarnClusterScheduler: Lost executor 396 on ip-172-32-78-221.ec2.internal: Container killed by YARN for exceeding memory limits. 5.5 GB of 5.5 GB physical memory used. Consider boosting spark.yarn.executor.memoryOverhead.
This happens for more than just this executor too; it looks like almost all of them.
I can try to increase the executor memory overhead, but I would like to know why getting the column names results in an OOM error. I wouldn't think that something that trivial would take up that much memory?
Update
I attempted to run the job with both spark.driver.memoryOverhead=7g and spark.yarn.executor.memoryOverhead=7g and I again got an OOM error