What is covered

• In this blog, I would describe my experiments about setting up a local (standalone) Spark cluster on Mac M1 machine.
• Would start jupyter notebook using anaconda navigator
• Submit jobs using Python code executing in the notebook with jobs submitted to the master created above.
• Explore the UI for:
  • Master
  • Worker
  • Jobs (transformation and action operations, Directed Acyclic Graph)

Why local (standalone) server

• local server is a simple deployment model and is possible to run the daemons in a single node
• Jobs submitted using pyspark are explored using the same UI in previous blog
• Jobs submitted using python programs are explored in this blog

Start the master and worker

• After installation of pyspark (using homebrew), scripts for starting the master and worker were available at the following location:

cd /opt/homebrew/Cellar/apache-spark/3.2.1/libexec/sbin

• start the server (note that would be specific to the machine)

% ./start-master.sh 
starting org.apache.spark.deploy.master.Master, logging to /opt/homebrew/Cellar/apache-spark/3.2.1/libexec/logs/<your-local-logfile>.out

• check the log file for the spark master server details ( is obtained from the previous step>

% tail -10 /opt/homebrew/Cellar/apache-spark/3.2.1/libexec/logs/<your-local-logfile>.out
...
22/05/18 07:41:57 INFO Utils: Successfully started service 'sparkMaster' on port 7077.
22/05/18 07:41:57 INFO Master: Starting Spark master at spark://<your-spark-master>:7077
22/05/18 07:41:57 INFO Master: Running Spark version 3.2.1
22/05/18 07:41:57 INFO Utils: Successfully started service 'MasterUI' on port 8080.
...

• check the master UI
• notice that workers are zero, since, we have not yet started any

• start the worker by selecting the number of cores and memory based on your system configuration.
• please check logfile in the previous step and replace with the appropriate value for

 % ./start-worker.sh --cores 2 --memory 2G spark://<your-spark-master>:7077
starting org.apache.spark.deploy.worker.Worker, logging to /opt/homebrew/Cellar/apache-spark/3.2.1/libexec/logs/<your-local-worker-logfile>.out

• check the master UI for worker information (now it shows 1 worker)

• also, check the worker UI for the cores and memory allocated

submit jobs using python in jupyter notebook and observe jobs UI

• started jupyter notebook using anaconda navigator
• write python code to submit jobs to the master setup above
• operations are performed on RDDs (Resilient Distributed Dataset)
• spark has two kinds of operations:
  • Transformation → operations such as map, filter, join or union that are performed on an RDD that yields a new RDD containing the result
  • Action → operations such as reduce, first, count that return a value after running a computation on an RDD

• Actions could be displayed as jobs in the UI and transformations could be observed while exploring the DAG (Directed Acyclyic Graph) output

• started jupyter notebook from anaconda navigator

• create Spark Application in notebook (replace with your-master)

import pyspark
from pyspark.sql import SparkSession
spark = SparkSession.builder.master("<your-master>") \
                    .appName('SparkTestPython') \
                    .getOrCreate()

print("First SparkContext:");
print("APP Name :"+spark.sparkContext.appName);
print("Master :"+spark.sparkContext.master);

• Now observe the master and worker that the Application is listed

• we will write a simple code in the notebook to do one transformation (parallelize) and one action (count) operation

  a=("hello","world","jupyter","local","mac","m1")
  b=spark.sparkContext.parallelize(a)
  b.count()
  

• From the master, click on the application to check the application UI:
• In the application UI, click on "Application Detail UI" to get details on the jobs

• to see DAG (Directed Acyclic Graph) visualization click on the job (under the Description column)
• this is a simple DAG with only one transformation and one operation

• we will write another simple code in the notebook to do two transformations (read from file and map) and one action (collect) operation

x=spark.sparkContext.textFile('/tmp/sparktest.txt')
y=x.map(lambda n: n.upper())
y.collect()

• Check the jobs UI

Next steps

• try out additional spark operations, and actions, and check additional details like storage, etc. in the UI
• using data frames

References

• Spark Standalone Server
• Anaconda Distribution