In this blog, I would like to discuss how you will be able to use Python to run a databricks notebook for multiple times in a parallel fashion. Noting that the whole purpose of a service like databricks is to execute code on multiple nodes called the workers in parallel fashion. But there are times where you need to implement your own parallelism logic to fit your needs.

To follow along, you need to have databricks workspace, create a databricks cluster and two notebooks. The parent notebook orchestrates the parallelism process and the child notebook will be executed in parallel fashion. The idea would be that the parent notebook will pass along a parameter for the child notebook and the child notebook will use that parameter and execute a given task. Without further to say, let’s get to it.

For simplicity let’s design a child notebook that takes a number as an input and then print the multiplication of this number by 10. Also, to make sure that we test our parallelism logic, we will introduce 20 seconds sleep time for our child notebook. So open up your child notebook and enter the following code in Command 1 (this code will help us pass a parameter from the parent notebook).

dbutils.widgets.text("numberToProcess","","")
dbutils.widgets.get("numberToProcess")
numberToProcess = int(getArgument("numberToProcess"))

Open up a new command in child notebook and enter the following code which will calculate the 10 multiplier for our number of interest, introduce a sleep time of 20 seconds and then print the output

import time
outputNumber = numberToProcess * 10
time.sleep(20) # sleep for 20 seconds 
print('The desired output is {}'.format(outputNumber))

So your child notebook will look something like this

Now that you have your child notebook setup, go to your parent notebook and paste the following code to import the multithreading packages

from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor

Then we define the function that will execute a child notebook passing in the number parameter

def processAnIntegerNumber(numberToProcess):
  dbutils.notebook.run(path = "/Shared/ChildNotebook",
                                        timeout_seconds = 300, 
                                        arguments = {"numberToProcess":numberToProcess})

Finally, the magic command, which will take a list of numbers, spin up a multithreading executor and then map the list of numbers to that executor. This command will execute the childnotebook instance for 5 times (because the list of numbers contains five numbers) in a parallel fashion.

listOfNumbers = [87,90,12,34,78]
with ThreadPoolExecutor() as executor:
  results = executor.map(processAnIntegerNumber, listOfNumbers)

Executing the parent notebook, you will notice that 5 databricks jobs will run concurrently each one of these jobs will execute the child notebook with one of the numbers in the list. This is a snapshot of the parent notebook after execution

Notice how the overall time to execute the five jobs is about 40 seconds. Now, if we open the link to any of these jobs, we will notice that the time was also about 32 seconds indicating that jobs were run in parallel

Hope this blog helps you run your jobs faster and satisfies your “Need for Speed”.

You can follow this blog for Machine learning, data engineering, devops, dynamics and power apps tips and tricks.