How to upload files to new EMR cluster

Question

I want to create a new EMR cluster, run a PySpark job and destroy it. Ideally I'd like to do this by adding a step when creating the cluster. The command I would run locally to start the jobs looks like this:

spark-submit calculate.py --input x.csv --output output

What I don't understand is how I can make sure that calculate.py is already available on the master node. I saw a reference of reading the python script from an S3 bucket here, but I couldn't get that to work.

Now I have separate commands for creating the cluster, putting the script on the master node and adding the steps. the problem with this is that the cluster keeps running after the job step finishes.

You have to use s3 or if you have your files in HDFS on a remote server, you can use that! Nevertheless, there is some plumbing to do to allow access. — eliasah, Sep 28 '15 at 16:33

score 7 · Accepted Answer · answered Sep 29 '15 at 12:46

I solved this by creating an extra step which just calls hadoop fs -copyToLocal to download the files.

I had the problem with the bootstrap step that the hadoop command was not installed yet.

Full working example using boto3:

import boto3

client = boto3.client('emr', region_name='eu-central-1')

cluster_id = client.run_job_flow(
    Name='My job',
    LogUri='s3://my-bucket/emr-logs/',
    ReleaseLabel='emr-4.0.0',
    Instances={
        'MasterInstanceType': 'm3.xlarge',
        'SlaveInstanceType': 'm3.xlarge',
        'InstanceCount': 2,
        'Ec2KeyName': 'my_key',
        'Ec2SubnetId': 'subnet-123456'
    },
    Applications=[
        {
            'Name': 'Spark'
        }
    ],
    Steps=[
        {
            'Name': 'Copy files to master',
            'ActionOnFailure': 'CONTINUE',
            'HadoopJarStep': {
                'Jar': 'command-runner.jar',
                'Args': [
                    'hadoop',
                    'fs',
                    '-copyToLocal',
                    '%s/code/*' % S3_BUCKET,
                    '/home/hadoop/'
                ]
            }
        },
        {
            'Name': 'Calculate step',
            'ActionOnFailure': 'CONTINUE',
            'HadoopJarStep': {
                'Jar': 'command-runner.jar',
                'Args': [
                    'spark-submit',
                    '/home/hadoop/calculate.py',
                    '--param',
                    'value'
                ]
            }
        }
    ],
    VisibleToAllUsers=True,
    JobFlowRole='EMR_EC2_DefaultRole',
    ServiceRole='EMR_DefaultRole')

print cluster_id

score 4 · Answer 2 · answered Sep 28 '15 at 19:26

One way is to copy the files you need onto the nodes in a bootstrap action. We have a small shell-script, s3-to-local.sh, located in a bucket on S3 that does exactly that. The script is trivial and looks something like this:

#!/bin/bash

echo "Copying ${1} to ${2}"
hadoop fs -copyToLocal ${1} ${2}

In my create-cluster command, I then add the bootstrap action:

aws --profile myProfile create-cluster \
--name "My cluster name" \
--auto-terminate \
--bootstrap-actions Path=s3://path/to/s3-to-local.sh,Name=copy-file,Args=[s3://path/to/file.ext,/local/filepath/file.ext] \
--steps ...

and the file.ext gets copied onto the nodes of my EMR cluster.

As for your cluster not terminating, you can add the auto-terminate flag like I have in the above. This will result in termination of your cluster when all steps have completed.

Note that there are other ways of doing it, but this is a simple and very straight forward way. And it works :)

How to upload files to new EMR cluster

2 Answers2