HTTPSConnectionPool(host=’files.pythonhosted.org’, port=443): Read timed out

I recently had an issue where one of our EMR clusters failed to bootstrap the python modules via PIP. I checked the logs and saw that we ran into the following error:

I wanted to have PIP not die if it timed out, I also wanted it to retry on failure. By adding the following to my bootstrap.sh I was able to have the PIP socket timeout at a longer interval, also bump up the retries to 10. I have not seen the issue since I applied the new settings.

From the PIP help page:

AttributeError: module ‘cryptography.utils’ has no attribute ‘register_interface’

I just recently came across an issue when we were bootstrapping one of our EMR clusters, looks like when trying to import pgpy we failed with the following traceback:

Apparently the cryptography team released a new version on September 7th 2022 that broke the pgpy library.
https://pypi.org/project/cryptography/38.0.1/

We needed to downgrade our version to get things working again. I figured I would post this to see if others run into this, according to the pgpy github page, they are working on a fix.

https://github.com/SecurityInnovation/PGPy/issues/402

Here is how I solved it in the meantime, I needed to downgrade the cryptography library.

AWS EMR ImportError: this version of pandas is incompatible with numpy < 1.17.3

I found another one that I thought was worth writing a quick blog post about. We use AWS Elastic Map Reduce with transient clusters, so in order to get the python libraries installed, we need to use the bootstrap feature. We ran into many issues trying the standard bootstrap script which looked something like this:

The contents of requirements.txt looked like this:

We would get all the nodes in the cluster to bootstrap properly however the logs showed the following:

And when trying to import from pyspark, we saw this:

After speaking with AWS support, it turns out this was a known issue. When a cluster is launched, EMR first provisions the EC2 instances, after that it runs the bootstrap actions. Thus, when the bootstrap action runs, it installs the desired version. However, since the applications are installed after the bootstrap action, these applications override the custom installation for the Python packages. In order to get around the issue of the version being overridden, the workaround is to make use of a Bootstrap Action that delays the installation of the packages until the nodes are fully up and running. This will resolve the conflict that we have been seeing with pandas and numpy. Here is what our final working bootstrap.sh looks like, hope this helps, it was a tough one to solve:

AWS Apache Managed Airflow EMR ModuleNotFoundError: No module named ‘requests’ Bootstrap

I came across another fun one the other day, we are in the process of migrating our on premise elastic map reduce system into the cloud. We are using AWS EMR and have AWS Managed Airflow as the executor (DAG). We came across an odd situation with a pyspark application. When using Airflow with a SparkSubmitHook, the job would bootstrap looking just fine according to the run logs, however it would fail with No module named 'requests' when the application tried to import it. This was very odd since we have this application running from spark-submit just fine when calling it from the master node command line.

I decided to investigate the differences, our bootstrap script for installing python modules via pip which we call from the EMR API RunJobFlow call looks like this:

This is very basic, all it does is upgrade PIP and run PIP install to install each of the modules. When checking the bootstrap log I can see that PIP upgrades and goes out to the repo and installs the packages just fine. So why were we getting the No module named 'requests' error when executing through airflow. After a ton of googling and research I have found the issue and applied a solution that worked. Turns out airflow will run as the root user when bootstrapping, so if you notice we use the --user argument in pip. This will instruct the packages to be installed in the calling users home directory, the kicker is the code is run by the hadoop user on the EMR cluster nodes after executing from airflow. So turns out, the hadoop user is unable to access the requests module since root installed it with --user. I changed the bootstrap script to the following and it all started working, by removing --user and prefixing with sudo, the packages now get installed in a globally available area for all users. I am sure there are better ways to do this, I am still learning and researching, but if you run into this, the change below with get you out of the woods.

After some further research, and testing we decided to utilize a requirements.txt file to be called by the bootstrap shell script in the RunJobFlow call, first create a requirements.txt file, I like to hardcode the versions so nothing changes unexpectedly as you bootstrap a new cluster and it reaches out to PyPy to get the packages.

https://docs.aws.amazon.com/emr/latest/APIReference/API_RunJobFlow.html

Add your desired packages and version numbers to a file called requirements.txt like below:

Then you will need to copy this file into a bucket you have access to:

Then create a shell script that has the following, call it bootstrap.sh:

Copy that shell script to your bucket:

And execute it via the bootstrap actions in the RunJobFlow EMR API call:

As you can see the shell script will be executed which will copy the requirements.txt file locally and then run pip -r against it which will install all the packages. If you want to see the log on a running cluster, you can ssh to the master node and view the logs here to see the bootstrapping take place:

You should see the stdout log as so:

Hope this helps.