DEV Community

Cover image for Provision Spark Clusters and Run Jobs Right in Your Notebook
Charles Landau
Charles Landau

Posted on • Originally published at charlesdlandau.io

Provision Spark Clusters and Run Jobs Right in Your Notebook

#python #tutorial

We'll be using GCP SDKs to provision Spark clusters and run PySpark jobs on them -- all from a notebook. In fact, the source for this document IS a notebook. You can clone this demo here:

https://github.com/CharlesDLandau/sparkAllInOne

This will take between five and ten minutes. I'm so spoiled by the public cloud that I don't even really understand how crazy this level of convenience is.

In this activity we will need a few prerequisites. They are:

  • A project directory with a python virtual environment. 
mkdir sparkDemo && cd sparkDemo
python -m virtualenv venv

// windows
call venv/Scripts/activate

// else
source venv/bin/activate

pip install google-cloud-dataproc jupyterlab
Enter fullscreen mode Exit fullscreen mode 
project_id = 'sparkdemo'
region = 'us-east1-b'
Enter fullscreen mode Exit fullscreen mode

  • A service account, and the JSON file with its credentials

    • Go to the IAM console and go to Service Accounts (left menu). Create a new account and fill in the details sensibly. Give the service account dataproc worker permissions. Remember to save the credentials file as JSON when prompted. My habit is to save it as client_secrets.json and then I immediately add that filename to my gitignore file.

With those in hand...lets set up our client 

from google.cloud import dataproc_v1
Enter fullscreen mode Exit fullscreen mode 
client = dataproc_v1.ClusterControllerClient.from_service_account_json(
    './client_secrets.json')
Enter fullscreen mode Exit fullscreen mode

Above, we imported the dataproc_v1 SDK, and we instantiated a client with the special from_service_account_json method.

# A parser for google's responses

def response_parser(client_obj, verbose=False):
    responses = []
    for page in client_obj.pages:
        for element in page:
            responses.append(element)

    if not responses and verbose:
        print('No responses found!')

    if responses and verbose:
        for resp in responses:
            print(resp)

    return responses

clusters = client.list_clusters(project_id, 'global')
responses = response_parser(clusters);
len(responses)

>> 0
Enter fullscreen mode Exit fullscreen mode

We are showing zero clusters provisioned. Let's provision a cluster.

## A response handler that we can attach to cluster actions

def handler(ops):
    result = ops.result()

    # Uncomment this to see results from the calls in this notebook
    # (this can also help you learn config options)
    # print(result)
Enter fullscreen mode Exit fullscreen mode

The important thing here is to create a dataproc_v1.types.Cluster instance.

config={'gce_cluster_config':{'zone_uri':region}}
cluster_config = dataproc_v1.types.Cluster(
    project_id=project_id, cluster_name="democluster", config=config)
Enter fullscreen mode Exit fullscreen mode

GCP uses sensible defaults to configure this instance, but you can use the printout from the callback handler to examine the config fields and override where appropriate. For example to scale up workers add the following to the config:

'worker_config':{'num_instances':3}
Enter fullscreen mode Exit fullscreen mode

Also note that the region is global at the call to create_cluster, and not in the gce_cluster_config.

creating = client.create_cluster(project_id, 'global', cluster_config)
creating.add_done_callback(handler)
Enter fullscreen mode Exit fullscreen mode

If you enabled printing in the callback, you can just wait for that to fire now. Once the cluster is up we can repeat our call to list_clusters and see that we have one now.

clusters = client.list_clusters(project_id, 'global')

responses = response_parser(clusters);
len(responses)

>> 1
Enter fullscreen mode Exit fullscreen mode

Hooray!

Run a job

Now let's set up a job controller.

jobber = dataproc_v1.JobControllerClient.from_service_account_json(
    './client_secrets.json'
)
Enter fullscreen mode Exit fullscreen mode

We also need somewhere to put our job files

from google.cloud import storage
import uuid

gs = storage.Client.from_service_account_json('./client_secrets.json')
bucket = gs.create_bucket(f'sparkdemo-{uuid.uuid1()}')
Enter fullscreen mode Exit fullscreen mode

We can store our code in the bucket. If you have Spark installed, this would be the part where you could start prototyping in local mode. For now, just copy the following code into a file main.py and save it in the root of your project.

import pyspark
import sys

sc = pyspark.SparkContext()

# Dummy task
terms = ["fodder" for x in range(9000)] + ["spartan", "sparring", "sparrow"]
words = sc.parallelize(terms)
words.filter(lambda w: w.startswith("spar")).take(2)

# Output target was passed to args config field
bucket = sys.argv[1]
output_directory = 'gs://{}/pyspark_output'.format(bucket)

# For csv output
sqlContext = pyspark.SQLContext(sc)
df = sqlContext.createDataFrame(words, pyspark.sql.types.StringType())
df.coalesce(1).write.format('com.databricks.spark.csv'
            ).options(header='true').save(output_directory)
Enter fullscreen mode Exit fullscreen mode

Here's a simple function that is barely modified from the boilerplate on the storage client library docs. We'll use it to upload files to our bucket, specifically our main.py file.

def upload_blob(bucket, source_file_name, destination_blob_name):
    """Uploads a file to the bucket."""
    blob = bucket.blob(destination_blob_name)

    blob.upload_from_filename(source_file_name)

    print('File {} uploaded to {}.'.format(
        source_file_name,
        destination_blob_name))

# Use the helper function to upload our code
upload_blob(bucket, './main.py', "input_files/main.py")

>> File ./main.py uploaded to input_files/main.py.
Enter fullscreen mode Exit fullscreen mode

Now that our code is in place, we need to make an instance of dataproc_v1.types.PySparkJob

pyspark_job = dataproc_v1.types.PySparkJob(
    main_python_file_uri=f"gs://{bucket.name}/input_files/main.py",
    args={bucket.name}
)
Enter fullscreen mode Exit fullscreen mode

main_python_file_uri points to our code. We could also pass python_file_uris and file_uris if we had more supporting files to send, but we don't.

Our main.py script uses sys.argv[1] to construct an output directory in our bucket, so we pass that to args.

We also need to place the job in our cluster:

placement = dataproc_v1.types.JobPlacement(cluster_name="democluster")
Enter fullscreen mode Exit fullscreen mode

We're ready to submit a job.

job = jobber.submit_job(project_id, 'global', {'placement': placement,
                                               'pyspark_job':pyspark_job})
Enter fullscreen mode Exit fullscreen mode

After a minute or so the job finishes. We can check on it like so:

job_status = jobber.get_job(project_id, 'global', job.job_uuid)
job_status.status

>>    state: RUNNING
>>    state_start_time {
>>      seconds: 1563547231
>>      nanos: 583000000
>>    }
Enter fullscreen mode Exit fullscreen mode

Try again...

job_status = jobber.get_job(project_id, 'global', job.job_uuid)
job_status.status

>>    state: DONE
>>    state_start_time {
>>      seconds: 1563547265
>>      nanos: 793000000
>>    }
Enter fullscreen mode Exit fullscreen mode

Jobs Done!

We did it!

Let's download contents from our bucket.

import os

try:
    os.mkdir('output')
except FileExistsError as e:
    pass

for page in bucket.list_blobs().pages:
    for element in page:
        print(element.name)
        blob = bucket.blob(element.name)
        destination = element.name.split('/')[-1]
        if destination:
            blob.download_to_filename(f"./output/{destination}")

>>    input_files/main.py
>>    pyspark_output/
>>    pyspark_output/_SUCCESS
>>    pyspark_output/part-00000-df4302c1-aec3-415b-84d6-66bd915ae938-c000.csv
Enter fullscreen mode Exit fullscreen mode

Our outputs are downloaded locally. We're done, so let's tear it all down.

destroying = client.delete_cluster(project_id, 'global', "democluster")
destroying.add_done_callback(handler)
Enter fullscreen mode Exit fullscreen mode 
# Clean and delete bucket
for page in bucket.list_blobs().pages:
    for element in page:
        bucket.delete_blob(element.name)
bucket.delete()
Enter fullscreen mode Exit fullscreen mode

There you have it. Now spark clusters are a disposable thing for you. When you want one, you can spin it up at any size your credit card will support. Prototype your jobs locally and submit them from your notebook. When you're done, tear it back down.

Why would you want to know how to do this? For one-off batch processing and reporting. When would you definitely not use this? In production.

Reference Docs

  1. https://googleapis.github.io/google-cloud-python/latest/dataproc/
  2. https://googleapis.github.io/google-cloud-python/latest/storage/
  3. https://spark.apache.org/docs/2.1.0/api/python/

Thank you!

Thank you for reading. I hope this has been helpful for you.

I'm always learning, so if you see anything wrong in here I hope you'll leave me a note to share your insights. Or just leave a note anyway.

Photo by Jez Timms on [Unsplash](https://unsplash.com/photos/_Ch_onWf38o)

Top comments (0)

Read next

matowang profile image

Never have unhandled Errors in TypeScript

Matthew Wang -

devstoriesplayground profile image

Top 10 Platforms to Practice Python

Devstories Playground -

miriacode profile image

How I built my first mechanical keyboard [Tutorial] [Part 2]

Miriam Acuña -

code42cate profile image

Day 13: Docker Multistage Builds

Jonas Scholz -