Scaling Data Science in Python with Dask

Using data to innovate, deliver, or disrupt in today’s marketplace requires the right software tools for the job. For decades, it has been difficult to process data that is large or complex on a single machine. Open source solutions have emerged to ease the processing of big data, and the Python and the PyData ecosystem offer an incredibly rich, mature and well-supported set of solutions for data science.

If you’re using Python and your team is like most, you are patching together disparate software and workloads, manually building and deploying images to run in containers, and creating ad hoc clusters. You may be working with Python packages like pandas or NumPy, which are extremely useful for data science but struggle with big data that exceeds the processing power of your machine.

Are you looking for a better way?

Dask is a free and open source library that provides a framework for distributed computing using Python syntax. With minimal code changes, you can process your data in parallel, which means it will take you less time to execute and give you more time for analysis.

We created this guide for data scientists, software engineers, and DevOps engineers. It will define parallel and distributed computing and the challenges of creating these environments for data science at scale. We will discuss the benefits of the Pythonic ecosystem and share how scientists and innovators use Dask to scale data science. From the terminology used to libraries available, it’s here.

We are Coiled, founded by Dask creators and core contributors. We hired open source engineers to maintain and improve Dask, so that it remains open while being stable and mature for enterprise needs.

We update this page regularly, so feel free to bookmark it for easy reference. If you see an opportunity to improve it, please share it with us.

Introduction: Will this guide be helpful to me?


This guide will be helpful to you if you are:

  • Using Python for data science, data analytics, and/or machine learning
  • Scaling your Python data science to larger data sets
  • Looking for a way to iterate and experiment faster so you can leverage a competitive advantage
  • Challenged to deploy a distributed computing system within an organization
  • Learning about best practices for distributed computing

The Basics: Distributed Computing and Python


What is parallel computing?

Parallel computing is a type of computation that allows many processes or calculations to be completed at the same time. Also known as parallel processing, parallel computing makes it easier to break large problems into smaller ones that can be solved at the same time. A parallel computing system consists of multiple processors that use shared memory to communicate with each other.

This time-lapse video of Amish workers raising a barn in 10 hours shows how people working in parallel can accomplish far more, far faster. Parallel computing establishes similar efficiencies of scale.
This time-lapse video of Amish workers raising a barn in 10 hours shows how people working in parallel can accomplish far more, far faster. Parallel computing establishes similar efficiencies of scale.

What is distributed computing?

Distributed computing is a model that connects a group of machines (i.e., computers) and where software can be shared among those machines. A distributed computing system consists of multiple processors, with their own memory, that are connected by a communication network. In this configuration, networked computers can communicate and coordinate actions by sending messages.

This communication is carried out by nodes, which are capable of creating, receiving, or transmitting information within the distributed system. While they run separately, each machine is part of a larger system. A cluster is the collection of nodes that is sharing compute power to complete processes.

While you can process data sequentially on a single computer, data science and machine learning models often require more data than your computer’s memory can handle. Sometimes, even when your processor can handle the data, it can only process it at a particular rate. Distributing your data over multiple computers reduces the amount of data held in memory by each one. Creating an architecture where more computers can work on a task at the same time means your task gets done faster.

 Distributed vs. Parallel Computing

This image shows the differences between a parallel computing system and a distributed computing system. A parallel computing system consists of multiple processors that use shared memory to communicate with each other. A distributed computing system consists of multiple machines, each with its own CPU(s) (central processing unit), that are connected by a communication network.
A parallel computing system consists of multiple processors that use shared memory to communicate with each other. A distributed computing system consists of multiple machines, each with its own CPU(s) (central processing unit), that are connected by a communication network.

What is scalable computing?

Scalable computing refers to a distributed architecture that expands beyond a single thread on a single core. It’s an approach that you can apply when you are working with larger datasets, larger workloads requiring more GPU or CPU, and more complexity.

What is Python?

Python is a programming language that is object-oriented, with dynamic semantics. Its principal author, Guido van Rossum, created it at Stichting Mathematisch Centrum in the Netherlands in the early 1990s as a successor to a language called ABC. Python includes contributions from many other people as well, and all Python releases are open source.

Python provides a well-regarded foundation for configuring components that serve your business best, rather than cobbling together bought components that may not be customized for your use case.

Why choose Python?

Python is a common choice for rapid application development, due to Its high-level, built-in data structures and dynamic typing and binding. Python is an excellent scripting or glue language that connects existing components (e.g., libraries). 

Python is low maintenance and has a syntax that makes it easy to read and understand, which reduces the cost of program maintenance. It’s easy to learn, and it supports a wide range of software packages and libraries that are readily available, which encourages program modularity and code reuse.

Programmers like Python for the increased productivity it can provide, with no compilation step and easy debugging. Data scientists like Python for its ease of use and ability to interface with high-performance algorithms.

Those who know Python sometimes use the term PyData or PyData stack to refer to the collection of commonly used Python libraries, software packages, and tools within the data science community.

Pythonic Ecosystem

This image shows the Pythonic ecosystem, with Python seated at the center bottom of the image. Upward and outward from Python are icons that represent popular libraries used with Python, including NumPy, Jupyter, Xarray, scikit-learn, and Dask. The image underscores that the Pythonic ecosystem is vast and expanding.
The Pythonic ecosystem is vast and expanding. Source: Jake VanderPlas, keynote, PyCon 2017

Python Libraries for Working With Data

There are Python software packages and libraries for working with various types of data. A number of Python programs are available in environments that are designed for specific industry applications. For example, PanGeo offers an ecosystem of compatible geoscience Python packages which follow well-established open source best practices.
You can install and manage Python packages from a repository of software. Conda is an open source package and environment management system for Python that runs on Microsoft Windows, Apple’s macOS, and Linux. You also can use pip to install packages from the Python Package Index (PyPI) and other indices.

Commonly used libraries for data science and machine learning include:

  • Dask is a Python-native open source library that provides a framework for distributed computing.
  • Jupyter is an open source project for interactive computing. It includes Jupyter Lab and Jupyter Notebooks that allow you to create and share documents that contain live code, equations, visualizations, and narrative text.
  • NumPy (Numerical Python) is a Python library used for working with arrays. It is an open source project that is free and available to use.
  • Pandas is a Python library used for data manipulation and analysis. It has data structures and operations for working with numerical tables and time series data. It is free and available to use.
  • Matplotlib is a Python library for making 2-D plots of arrays. It uses NumPy and other extension code that ensures high performance, even for large arrays.
  • Scikit-learn is a Python library with unsupervised and supervised machine learning algorithms.
  • SciPy (Scientific Python) is a scientific computation library that uses NumPy. It provides utility functions for optimization, statistics, and signal processing.
  • Seaborn is a Python library based on Matplotlib. It provides a high-level interface for data visualization and for drawing statistical graphics.
  • Xarray is a Python library built on NumPy with labeled dimensions. It’s popular for use with geospatial data. 
  • XGBoost is a Python library with gradient boosted decision trees designed for speed and performance.

Use Cases for Distributed Computing and Python

Python, along with the libraries in the growing PyData ecosystem, power some of the most exciting research happening today:

  • LIGO Lab (short for Caltech and MIT’s Laser Interferometer Gravitational-Wave Observatory) used operations support system (OSS) tools from the PyData ecosystem in its discovery of gravitational waves in 2015.
  • In 2018, Netflix was running more than 150,000 Jupyter notebook jobs against data collected by a streaming pipeline of more than one trillion events every day so they could deliver personalized experiences for 130 million members around the world. Netflix later developed an internal notebook, Polynote, which they released via open source to the public in 2019.
  • The team that developed the Event Horizon telescope (EHT) used the PyData stack to create the first-ever image of a black hole. EHT is an array of eight ground-based radio telescopes forming a virtual computational telescope that allows scientists to study the universe with unprecedented sensitivity and resolution.

These domains derive scientific and business value from massive amounts of data. Their stories show what is possible when community-based, open source software is used to understand and solve some of the world’s most challenging problems.

The Pythonic ecosystem is evolving and expanding. As with the Netflix example, technology companies using data to face big challenges often develop and release great tools that are adopted by the industry.

What is Spark?

Apache Spark is a distributed computing tool for tabular datasets that is popular for big data analysis. Spark has high-level APIs in Java, Scala, Python, and R. Spark has inspired its own ecosystem, and integrates well with many other Apache projects. If you prefer Scala or the SQL language, use mostly JVM infrastructure and legacy systems, Spark is a good all-in-one solution for business analytics and lightweight machine learning.

What is Dask?


Dask is a free and open source library that provides a framework for distributed computing in Python. Dask offers an advanced parallel computing environment for analytics, with performance at scale. It’s Python native, it supports all of the software tools and libraries you use with minimal code changes, and makes those tools screaming fast.

With Dask, you can:

  1. Process out-of-core computation on a single machine when data won’t fit in memory.
  2. Process data in parallel on a single machine using the same tools you already use.
  3. Process your data in parallel in a distributed architecture. This architecture could be on-prem or it can leverage cloud services, such as AWS (Amazon Web Services), Microsoft Azure, and GCP (Google Cloud Platform).

Dask provides a consistent API (application programming interface) for Python software packages that are familiar to data scientists and leverage the same underlying data structures. It can be used with any Python code and used alongside other software. Dask makes it seamless to switch between NumPy, pandas, and scikit-learn to their Dask-powered equivalents.

This architecture is particularly seamless for pandas because Dask DataFrame is a drop-in replacement for pandas DataFrames. The Dask DataFrame is a large parallel data structure composed of many smaller pandas DataFrames. One Dask DataFrame operation triggers operations on the component pandas DataFrames.


It is worth mentioning that there are a growing number of options to automate building, running, and monitoring your data workflows and pipelines. Among them are Apache Airflow and Prefect Cloud. Prefect works particularly well with Dask.

Benefits of Dask

There are many reasons data scientists and machine learning engineers use Dask. Here are a few of the benefits of distributed computing with Dask:

  1. Dask works with all of your favorite tools in the Pythonic ecosystem.
  2. Dask provides an easy-to-use framework for parallel computations across multiple cores, on a single workstation, or across multiple nodes in a cluster.
  3. Dask can scale from a single node to thousand-node clusters, making it ideal for large-scale data analysis. You can scale the same computations you would typically process in-memory to larger tasks using Dask.
  4. Dask accelerates your existing workflow with little to no code changes for data science that is screaming fast.

How does Dask work?

Dask provides parallelism to expand and speed processing, which is especially helpful for practicing data science and machine learning. Dask collections provide the API used to write Dask code. Collections create task graphs that define how to perform the computation in parallel.

In Dask, the computation is performed on a cluster, which consists of:

  • A client, which is the user-facing entry point where you write your Python code;
  • A scheduler, which receives tasks from the client and manages the flow of work and sends tasks to other machines (the workers); and
  • Workers, which compute the tasks the scheduler assigns to them.
This is an image of a Dask cluster. At the top is one box representing the client. The client is the user-facing entry point where you write your Python code. Under the client is another box representing the scheduler, which receives tasks from the client, manages the flow of work, and sends tasks to other machines (the workers). Under the scheduler are three boxes representing the workers, which compute the tasks and store and serve computed results to other workers or clients. The scheduler and the workers are the Dask distributed cluster. Source: dask.org
In Dask, the client is the user-facing entry point where you write your Python code. The scheduler receives tasks from the client, manages the flow of work, and sends tasks to other machines (the workers). The workers compute the tasks and store and serve computed results to other workers or clients. Source: dask.org
On the left, this image shows types of Dask collections: Dask Array, Dask DataFrame, Dask Bag, Dask Delayed, and Futures. In the middle, the image shows a task graph, with circles and squares representing the tasks to be done, in a particular workflow. On the right, the image shows the schedulers, which can be on a single machine or distributed. Collections provide the API used to write Dask code. Collections create task graphs that define how to perform the computation in parallel. The computation is performed on a cluster, which can be a single machine or a distributed environment. Source: dask.org
Dask collections provide the API used to write Dask code. Collections create task graphs that define how to perform the computation in parallel. The computation is performed on a cluster, which can be a single machine or a distributed environment. Source: Dask

Dask dashboards provide a live, interactive visual with multiple plots and tables to show the state of your cluster and ease diagnosis of problems. Some of these dashboards include:

  • Graph that maps tasks by those that are released, in memory, processing, or waiting;
  • Cluster map to visualize interactions between the scheduler and workers;
  • Task stream to show real-time activities performed by each worker;
  • Progress bar to track the progression of each task; and
  • Workers to show CPU and memory use by machine.
Dask provides a dashboard to view and diagnose the state of your cluster, workers, tasks, and progress. This image shows a screenshot of a dashboard in Dask. The dashboard has five modules with these names: Dask Graph, Dask Cluster Map, Dask Task Stream, Dask Progress, and Dask Workers. Source: Coiled.io
Dask provides a dashboard to view and diagnose the state of your cluster, workers, tasks, and progress. Source: Coiled.io

Who Uses Dask?


Dask supports a variety of workloads and use cases in data science and machine learning. Data scientists, researchers, software engineers, and DevOps engineers use Dask in business, academic, and government organizations to solve complex problems at massive scale.

Dask for the Enterprise

Enterprise organizations use Dask and the Pythonic ecosystem to process massive data so they can accelerate analysis, innovate products, discover new scientific perspectives, and serve customers better.

  • Walmart uses Python, Dask, and XGBoost for product forecasting of more than 500 million store-item combinations each week. They use machine learning algorithms to solve tough problems across the business, including supply chain management and last-mile delivery, so they can provide better service to more than 265 million customers each week.
  • NASA (National Aeronautics and Space Administration) is using Dask and the PyData accelerate data analysis. They developed an advanced Python application programming interface (API) with Xarray and Dask and integrated it with the Pangeo geoscience ecosystem to apply scientific tools for preprocessing, regridding, machine learning, and visualization.
  • Harvard Medical School researchers used Dask and napari, a multi-dimensional image viewer for Python, to preprocess and view massive image datasets interactively. This capability allowed them to view the data in unique ways; for example, they can view the process of a cell undergoing mitosis. Dask makes it easy to process massive image datasets interactively at scale.
  • Capital One uses Dask to scale Python workloads and reduce model training times. Early implementations of Dask reduced training times by 91% within a few months of development effort. Capital One’s team chose Dask because it offers a familiar API, so its developers would spend less time learning new software or adjusting their code for compatibility.

Local Dask Cluster Performance in a Model Training Pipeline

This image shows a graph mapping the improvements the Capital One team made with a local Dask cluster during one step in a model training pipeline. Their old process took them nearly 4,000 seconds to perform, whereas the local Dask cluster took nearly 1,000 seconds. For this team, Dask speeds iteration cycles during development, which allows developers to test their code faster. Source: Capital One
This figure shows improvements the Capital One team made with a local Dask cluster during one step in a model training pipeline. For this team, Dask speeds iteration cycles during development, which allows developers to test their code faster. (Source: Capital One)

Coiled: Managed Dask in the Cloud


Unlocking the power of PyData at scale first requires solving DevOps challenges, most of which are outside the experience of most data science and ML practitioners.

To practice data science at scale, you will need to provision machines on the cloud or on-prem, set up Kubernetes, authenticate users and apply quotas, manage custom software environments and rapidly changing docker images, and ensure secure networking and data access. You’ll also have to keep a system running around the clock with limited staff.

The creators of Dask created Coiled Cloud to make it easy to scale your existing Python workflows and collaboration-ready Dask projects across diverse use cases.

Coiled provides:

  • Managed software environments – Coiled Cloud builds managed software environments, launching clusters in the cloud that support the exact stack you need. Using the Conda and Pip environment files you already use, Coiled quickly builds and stores docker images to your exact specification.
  • GPU support to ease changes in hardware architectures and explore newly accelerated libraries like XGBoost, RAPIDS, PyTorch, and more
  • Global multi-region support
  • Cost management features, including user quotas, GPU access monitoring, and default idle timeouts
  • End-to-end network security
  • Run from anywhere – Coiled offers a web interface, but many people use Coiled from their own Python environment. Using Coiled Cloud, you can run from any Python script, including other web services, automated jobs, and from an interactive Jupyter session on your laptop.

Coiled Cluster Table

This image shows a Coiled cluster table, which shows critical details about that cluster, including status (i.e., running or stopped), number of workers, configuration, and cost. Source: Coiled.io
A Coiled cluster table provides visibility into critical cluster details, such as status, number of workers, configuration, and cost. Source: Coiled.io

Accelerate your data science and machine learning projects today on a Coiled cluster. Coiled can handle security, container environments, and team management, so you can get back to doing data science. Get started for free today on Coiled Cloud.

Are you ready to accelerate your data science workflow?

Frequently Googled Questions


Dask is an open source library that provides a framework for distributed computing in Python.

Dask offers an advanced parallel computing environment for analytics, with performance at scale for the software tools and Python libraries data scientists commonly use.

With Dask, you can process your data sequentially, by reading from a hard disk, on a single computer. You also can process your data in parallel on multiple computers in a distributed architecture that leverages cloud services, such as AWS (Amazon Web Services), Microsoft Azure, and GCP (Google Cloud Platform). Dask enables both approaches.

When you are working with pandas, a Dask DataFrame makes it possible to parallelize your infrastructure. A Dask DataFrame is a large parallel data structure that is made up of many smaller pandas DataFrames arranged along the index. These DataFrames may be stored on disk for computing on a single machine or in a cluster, on many different machines for distributed computing. One Dask DataFrame operation triggers many operations on the component pandas DataFrames.

Dask accelerates processing in pandas. A Dask DataFrame is a large parallel data structure that is made up of many smaller pandas DataFrames arranged along the index.

Python is great for data science at scale. Python, along with the libraries in its growing ecosystem, power some of the most exciting research happening today. Dask is a free and open source library that provides a framework for distributed computing in Python. Its advanced environment for analytics makes it easy to burst to the cloud as needed, when you need more memory, compute, or processing power. This is especially helpful when you are working with large datasets.

Dask provides advanced parallelism to expand and speed processing, which is especially helpful for practicing data science at scale. Dask collections provide the API used to write Dask code. Collections create task graphs that define how to perform the computation in parallel.

In Dask, the computation is performed on a cluster, which consists of:

  • A client, which is the user-facing entry point where you write your Python code;
  • A scheduler, which receives tasks from the client and manages the flow of work and sends tasks to other machines (the workers); and
  • Workers, which compute the tasks the scheduler assigns to them.

Dask is helpful when your models require more data than your computer’s memory can handle and you want to work within the Python ecosystem of software tools and libraries that are familiar to you. Dask makes it easy to use more processors, or more computers, to effectively work on your task at the same time and get it done faster.

Data scientists typically use Dask to:

  • Scale big-data analysis and process computations faster using Python
  • Process parallel computations across multiple cores, either on a single workstation, or across multiple nodes in a cluster
  • Scale data analysis without major changes to the data pipeline

For example, if you have a large dataset (80GB), your laptop’s memory isn’t vast enough to load it, so you might chunk the data by hand to analyze it. Dask accelerates and expands your compute power by automating the setup of new Python processes (i.e., workers) on your computer to call on for help. Each of these processes runs independently in a cluster.

A Dask DataFrame is a large parallel data structure that is made up of many smaller pandas DataFrames arranged along the index. These DataFrames may be stored on disk for computing on a single machine or on many different machines for distributed computing, in a cluster. One Dask DataFrame operation triggers many operations on the component pandas DataFrames.

In general, Dask is similar to Spark in the functionality it provides. Dask is more appropriate if you prefer Python-native code and/or have large legacy code bases you do not want to rewrite entirely. Dask is integrated into NumPy and pandas, making Dask much easier than Spark for users of those libraries to learn from scratch. Dask can be better if your use case is complex or does not fit the Apache Spark computing model. If you already use the Apache ecosystem, it might make sense to use Spark for your data science projects. See Dask’s comparison to Spark.