I typically install packages in EMR through Spark's install_pypi_package method. This limits where I can install packages from. How can I install a package from a specific GitHub branch? Is there a way I can do this through the install_pypi_package method?

CodePudding user response：

If you have access to cluster creation step, you can install the package using pip from github at bootstrap. (install_pypi_package is needed because the cluster is already running at that time and packages might not resolve on all nodes)

Installing prior Cluster is running:

A simple example (e.g with download.sh bootstrap file) of bootstrap and installing from github using pip is

#!/bin/bash
sudo pip install <you-repo>.git 

then you can use this bash at bootstrap as

aws emr create-cluster --name "Test cluster" --bootstrap-actions Path="s3://elasticmapreduce/bootstrap-actions/download.sh"

or you can use pip3 in bootstrap

sudo pip3 install <you-repo>.git 

or just clone it and build it locally on EMR with setup.py file

#!/bin/bash
git clone <your-repo>.git
sudo python setup.py install

After Cluster is running (Complex and not recommended)

If you still want to install or build a custom package when the cluster is already running, AWS has some explanation here that uses AWS-RunShellScript to install package on all core nodes. It says

(I) Install the package to Master node, (doing pip install on running cluster via shell or a jupyter notebook on top of it)

(II) Running following script locally on EMR, for which you pass cluster-id and boostrap script path(for e.g download.sh above) as arguments.

import argparse
import time
import boto3


def install_libraries_on_core_nodes(
        cluster_id, script_path, emr_client, ssm_client):
    """
    Copies and runs a shell script on the core nodes in the cluster.

    :param cluster_id: The ID of the cluster.
    :param script_path: The path to the script, typically an Amazon S3 object URL.
    :param emr_client: The Boto3 Amazon EMR client.
    :param ssm_client: The Boto3 AWS Systems Manager client.
    """
    core_nodes = emr_client.list_instances(
        ClusterId=cluster_id, InstanceGroupTypes=['CORE'])['Instances']
    core_instance_ids = [node['Ec2InstanceId'] for node in core_nodes]
    print(f"Found core instances: {core_instance_ids}.")

    commands = [
        # Copy the shell script from Amazon S3 to each node instance.
        f"aws s3 cp {script_path} /home/hadoop",
        # Run the shell script to install libraries on each node instance.
        "bash /home/hadoop/install_libraries.sh"]
    for command in commands:
        print(f"Sending '{command}' to core instances...")
        command_id = ssm_client.send_command(
            InstanceIds=core_instance_ids,
            DocumentName='AWS-RunShellScript',
            Parameters={"commands": [command]},
            TimeoutSeconds=3600)['Command']['CommandId']
        while True:
            # Verify the previous step succeeded before running the next step.
            cmd_result = ssm_client.list_commands(
                CommandId=command_id)['Commands'][0]
            if cmd_result['StatusDetails'] == 'Success':
                print(f"Command succeeded.")
                break
            elif cmd_result['StatusDetails'] in ['Pending', 'InProgress']:
                print(f"Command status is {cmd_result['StatusDetails']}, waiting...")
                time.sleep(10)
            else:
                print(f"Command status is {cmd_result['StatusDetails']}, quitting.")
                raise RuntimeError(
                    f"Command {command} failed to run. "
                    f"Details: {cmd_result['StatusDetails']}")


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('cluster_id', help="The ID of the cluster.")
    parser.add_argument('script_path', help="The path to the script in Amazon S3.")
    args = parser.parse_args()

    emr_client = boto3.client('emr')
    ssm_client = boto3.client('ssm')

    install_libraries_on_core_nodes(
        args.cluster_id, args.script_path, emr_client, ssm_client)


if __name__ == '__main__':
    main()