AWS Glue executor memory limit

Question

I found that AWS Glue set up executor's instance with memory limit to 5 Gb --conf spark.executor.memory=5g and some times, on a big datasets it fails with java.lang.OutOfMemoryError. The same is for driver instance --spark.driver.memory=5g. Is there any option to increase this value?

Solution 1

The official glue documentation suggests that glue doesn't support custom spark config.

There are also several argument names used by AWS Glue internally that you should never set:

--conf — Internal to AWS Glue. Do not set!

--debug — Internal to AWS Glue. Do not set!

--mode — Internal to AWS Glue. Do not set!

--JOB_NAME — Internal to AWS Glue. Do not set!

Any better suggestion on solving this problem?

Solution 2

despite aws documentation stating that the --conf parameter should not be passed, our AWS support team told us to pass --conf spark.driver.memory=10g which corrected the issue we were having

Solution 3

You can override the parameters by editing the job and adding job parameters. The key and value I used are here:

Key: --conf

Value: spark.yarn.executor.memoryOverhead=7g

This seemed counterintuitive since the setting key is actually in the value, but it was recognized. So if you're attempting to set spark.yarn.executor.memory the following parameter would be appropriate:

Key: --conf

Value: spark.yarn.executor.memory=7g

Solution 4

1. Open Glue> Jobs > Edit your Job> Script libraries and job parameters (optional) > Job parameters near the bottom
2. Set the following: key: --conf value: spark.yarn.executor.memoryOverhead=1024 spark.driver.memory=10g

Solution 5

I hit out of memory errors like this when I had a highly skewed dataset. In my case, I had a bucket of json files that contained dynamic payloads that were different based on the event type indicated in the json. I kept hitting Out of Memory errors no matter if I used the configuration flags indicated here and increased the DPUs. It turns out that my events were highly skewed to a couple of the event types being > 90% of the total data set. Once I added a "salt" to the event types and broke up the highly skewed data I did not hit any out of memory errors.

Here's a blog post for AWS EMR that talks about the same Out of Memory error with highly skewed data. https://medium.com/thron-tech/optimising-spark-rdd-pipelines-679b41362a8a

View more solutions

Comments

[Alexey Bakulin]

Thanks Kris. I will test your solution as soon as I can.

[Xavi]

I just added the following in my job section on my CloudFormation template, in the DefaultArguments part: "--conf": "spark.yarn.executor.memory=8g" without luck. The job fails with the message Container killed by YARN for exceeding memory limits. 5.7 GB of 5.5 GB physical memory used. I can actually see the parameter in the Job Parameters.

[Dwarrior]

I tried following setting with key as --conf and value as spark.driver.extraClassPath=s3://temp/jsch-0.1.55.jar for giving precedence to latest jar of jsch instead of the version that Glue is selecting but it doesn't work. Am I missing something. Also, as @rileyss mentioned, Glue documentation states that conf cannot be set. So, how should we go about resolving this?

[Dwarrior]

Have you been able to figure out the resolution for this? I tried following setting with key as --conf and value as spark.driver.extraClassPath=s3://temp/jsch-0.1.55.jar for giving precedence to latest jar of jsch instead of the version that Glue is selecting but it doesn't work. Am I missing something? So, how should we go about resolving this?

[cozyss]

@Dwarrior I'm not sure if you can customize anything about spark on Glue. It seems that Glue runs on a pre-set environment and that's why it's cheap. My solution is dividing the input data into smaller chunks and run several glue jobs. If you really need to use customized spark settings, you can try AWS EMR, which gives you much more freedom in adjusting spark parameters.

[Dwarrior]

thanks! Will explore the other options. I fathomed from other answers that some settings did work. :)

[selle]

@Xavi It could very well be the driver's config you need to modify. E.g "spark.driver.memory=8g"