Ben G Weber - Data Science in Production - Building Scalable Model Pipelines With Python-Independently Published (2020)

Ben G.
Weber
Data Science in Production:

Building Scalable Model
Pipelines with Python
Contents
Preface vii
0.1 Prerequisites . . . . . . . . . . . . . . . . . . . . vii
0.2 Book Contents . . . . . . . . . . . . . . . . . . . viii
0.3 Code Examples . . . . . . . . . . . . . . . . . . . x
0.4 Acknowledgements . . . . . . . . . . . . . . . . . x
1 Introduction 1
1.1 Applied Data Science . . . . . . . . . . . . . . . 3
1.2 Python for Scalable Compute . . . . . . . . . . . 4
1.3 Cloud Environments . . . . . . . . . . . . . . . . 6
1.3.1 Amazon Web Services (AWS) . . . . . . . 7
1.3.2 Google Cloud Platform (GCP) . . . . . . 8
1.4 Coding Environments . . . . . . . . . . . . . . . 9
1.4.1 Jupyter on EC2 . . . . . . . . . . . . . . . 9
1.5 Datasets . . . . . . . . . . . . . . . . . . . . . . 13
1.5.1 BigQuery to Pandas . . . . . . . . . . . . 15
1.5.2 Kaggle to Pandas . . . . . . . . . . . . . . 18
1.6 Prototype Models . . . . . . . . . . . . . . . . . 19
1.6.1 Linear Regression . . . . . . . . . . . . . . 20
1.6.2 Logistic Regression . . . . . . . . . . . . . 21
1.6.3 Keras Regression . . . . . . . . . . . . . . 22
1.7 Automated Feature Engineering . . . . . . . . . 26
1.8 Conclusion . . . . . . . . . . . . . . . . . . . . . 31
2 Models as Web Endpoints 33

2.1 Web Services . . . . . . . . . . . . . . . . . . . . 34
2.1.1 Echo Service . . . . . . . . . . . . . . . . 35
2.2 Model Persistence . . . . . . . . . . . . . . . . . 40
2.2.1 Scikit-Learn . . . . . . . . . . . . . . . . . 40
2.2.2 Keras . . . . . . . . . . . . . . . . . . . . 42
iii
iv Contents
2.3 Model Endpoints . . . . . . . . . . . . . . . . . . 43

2.3.1 Scikit-Learn . . . . . . . . . . . . . . . . . 44
2.3.2 Keras . . . . . . . . . . . . . . . . . . . . 46
2.4 Deploying a Web Endpoint . . . . . . . . . . . . 48
2.4.1 Gunicorn . . . . . . . . . . . . . . . . . . 48
2.4.2 Heroku . . . . . . . . . . . . . . . . . . . 49
2.5 Interactive Web Services . . . . . . . . . . . . . . 52
2.5.1 Dash . . . . . . . . . . . . . . . . . . . . . 52
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . 56
3 Models as Serverless Functions 57

3.1 Managed Services . . . . . . . . . . . . . . . . . 58
3.2 Cloud Functions (GCP) . . . . . . . . . . . . . . 59
3.2.1 Echo Service . . . . . . . . . . . . . . . . 60
3.2.2 Cloud Storage (GCS) . . . . . . . . . . . . 64
3.2.3 Model Function . . . . . . . . . . . . . . . 66
3.2.4 Keras Model . . . . . . . . . . . . . . . . 70
3.2.5 Access Control . . . . . . . . . . . . . . . 72
3.2.6 Model Refreshes . . . . . . . . . . . . . . 73
3.3 Lambda Functions (AWS) . . . . . . . . . . . . . 74
3.3.1 Echo Function . . . . . . . . . . . . . . . 74
3.3.2 Simple Storage Service (S3) . . . . . . . . 76
3.3.3 Model Function . . . . . . . . . . . . . . . 78
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . 85
4 Containers for Reproducible Models 87

4.1 Docker . . . . . . . . . . . . . . . . . . . . . . . 88
4.2 Orchestration . . . . . . . . . . . . . . . . . . . . 92
4.2.1 AWS Container Registry (ECR) . . . . . . 93
4.2.2 AWS Container Service (ECS) . . . . . . . 97
4.2.3 Load Balancing . . . . . . . . . . . . . . . 102
4.3 Kubernetes on GCP . . . . . . . . . . . . . . . . 104
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . 107
5 Workflow Tools for Model Pipelines 109

5.1 Sklearn Workflow . . . . . . . . . . . . . . . . . 110
5.2 Cron . . . . . . . . . . . . . . . . . . . . . . . . 116
5.2.1 Cloud Cron . . . . . . . . . . . . . . . . . 118
Contents v
5.3 Workflow Tools . . . . . . . . . . . . . . . . . . . 120

5.3.1 Apache Airflow . . . . . . . . . . . . . . . 121
5.3.2 Managed Airflow . . . . . . . . . . . . . . 125
5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . 127
6 PySpark for Batch Pipelines 129

6.1 Spark Environments . . . . . . . . . . . . . . . . 131
6.1.1 Spark Clusters . . . . . . . . . . . . . . . 132
6.1.2 Databricks Community Edition . . . . . . 133
6.2 Staging Data . . . . . . . . . . . . . . . . . . . . 136
6.2.1 S3 Credentials . . . . . . . . . . . . . . . 137
6.3 A PySpark Primer . . . . . . . . . . . . . . . . . 139
6.3.1 Persisting Dataframes . . . . . . . . . . . 140
6.3.2 Converting Dataframes . . . . . . . . . . . 143
6.3.3 Transforming Data . . . . . . . . . . . . . 145
6.3.4 Pandas UDFs . . . . . . . . . . . . . . . . 150
6.3.5 Best Practices . . . . . . . . . . . . . . . . 154
6.4 MLlib Batch Pipeline . . . . . . . . . . . . . . . 155
6.4.1 Vector Columns . . . . . . . . . . . . . . . 157
6.4.2 Model Application . . . . . . . . . . . . . 157
6.5 Distributed Deep Learning . . . . . . . . . . . . 161
6.5.1 Model Training . . . . . . . . . . . . . . . 162
6.5.2 Model Application . . . . . . . . . . . . . 163
6.6 Distributed Feature Engineering . . . . . . . . . 166
6.6.1 Feature Generation . . . . . . . . . . . . . 167
6.6.2 Feature Application . . . . . . . . . . . . 169
6.7 GCP Model Pipeline . . . . . . . . . . . . . . . . 171
6.7.1 BigQuery Export . . . . . . . . . . . . . . 171
6.7.2 GCP Credentials . . . . . . . . . . . . . . 172
6.7.3 Model Pipeline . . . . . . . . . . . . . . . 174
6.8 Productizing PySpark . . . . . . . . . . . . . . . 178
6.9 Conclusion . . . . . . . . . . . . . . . . . . . . . 179
7 Cloud Dataflow for Batch Modeling 181

7.1 Apache Beam . . . . . . . . . . . . . . . . . . . 183
7.2 Batch Model Pipeline . . . . . . . . . . . . . . . 188
7.2.1 Model Training . . . . . . . . . . . . . . . 188
vi Contents
7.2.2 BigQuery Publish . . . . . . . . . . . . . . 190

7.2.3 Datastore Publish . . . . . . . . . . . . . 196
7.3 Conclusion . . . . . . . . . . . . . . . . . . . . . 199
8 Streaming Model Workflows 201

8.1 Spark Streaming . . . . . . . . . . . . . . . . . . 202
8.1.1 Apache Kafka . . . . . . . . . . . . . . . . 203
8.1.2 Sklearn Streaming . . . . . . . . . . . . . 206
8.2 Dataflow Streaming . . . . . . . . . . . . . . . . 213
8.2.1 PubSub . . . . . . . . . . . . . . . . . . . 214
8.2.2 Natality Streaming . . . . . . . . . . . . . 216
8.3 Conclusion . . . . . . . . . . . . . . . . . . . . . 221
8.4 Thank You . . . . . . . . . . . . . . . . . . . . . 222
Preface
This book was developed using the leanpub1 platform. Please send
any feedback or corrections to: [email protected]
The data science landscape is constantly evolving, because new
tools and libraries are enabling smaller teams to deliver more im-
pactful products. In the current state, data scientists are expected
to build systems that not only scale to a single product, but a
portfolio of products. The goal of this book is to provide data
scientists with a set of tools that can be used to build predictive
model services for product teams.
This text is meant to be a Data Science 201 course for data science
practitioners that want to develop skills for the applied science dis-
cipline. The target audience is readers with past experience with
Python and scikit-learn than want to learn how to build data prod-
ucts. The goal is to get readers hands-on with a number of tools
and cloud environments that they would use in industry settings.
0.1 Prerequisites
This book assumes that readers have prior knowledge of Python
and Pandas, as well as some experience with modeling packages
such as scikit-learn. This is a book that will focus on breadth
rather than depth, where the goal is to get readers hands on with
a number of different tools.
Python has a large library of books available, covering the lan-
guage fundamentals, specific packages, and disciplines such as data
1
https://leanpub.com/ProductionDataScience
vii
viii Preface
science. Here are some of the books I would recommend for readers
to build additional knowledge of the Python ecosystem.
• Python And Pandas
– Data Science from Scratch (Grus, 2015): Introduces
Python from a data science perspective.
– Python for Data Analysis (McKinney, 2017): Provides ex-
tensive details on the Pandas library.
• Machine Learning
– Hands-On Machine Learning (Géron, 2017): Covers scikit-
learn in depth as well as TensorFlow and Keras.
– Deep Learning for Python (Chollet, 2017): Provides an ex-
cellent introduction to deep learning concepts using Keras
as the core framework.
I will walk through the code samples in this book in detail, but
will not cover the fundamentals of Python. Readers may find it
useful to first explore these texts before digging into building large
scale pipelines in Python.
0.2 Book Contents

The general theme of the book is to take simple machine learning
models and to scale them up in different configurations across mul-
tiple cloud environments. Here’s the topics covered in this book:
1. Introduction: This chapter will motivate the use of

Python and discuss the discipline of applied data science,
present the data sets, models, and cloud environments
used throughout the book, and provide an overview of
automated feature engineering.
2. Models as Web Endpoints: This chapter shows how
to use web endpoints for consuming data and hosting ma-
chine learning models as endpoints using the Flask and
Gunicorn libraries. We’ll start with scikit-learn models
and also set up a deep learning endpoint with Keras.
Book Contents ix
3. Models as Serverless Functions: This chapter will

build upon the previous chapter and show how to set
up model endpoints as serverless functions using AWS
Lambda and GCP Cloud Functions.
4. Containers for Reproducible Models: This chapter
will show how to use containers for deploying models with
Docker. We’ll also explore scaling up with ECS and Ku-
bernetes, and building web applications with Plotly Dash.
5. Workflow Tools for Model Pipelines: This chapter
focuses on scheduling automated workflows using Apache
Airflow. We’ll set up a model that pulls data from Big-
Query, applies a model, and saves the results.
6. PySpark for Batch Modeling: This chapter will intro-
duce readers to PySpark using the community edition of
Databricks. We’ll build a batch model pipeline that pulls
data from a data lake, generates features, applies a model,
and stores the results to a No SQL database.
7. Cloud Dataflow for Batch Modeling: This chapter
will introduce the core components of Cloud Dataflow and
implement a batch model pipeline for reading data from
BigQuery, applying an ML model, and saving the results
to Cloud Datastore.
8. Streaming Model Workflows: This chapter will intro-
duce readers to Kafka and PubSub for streaming mes-
sages in a cloud environment. After working through this
material, readers will learn how to use these message bro-
kers to creating streaming model pipelines with PySpark
and Dataflow that provide near real-time predictions.
After working through this material, readers should have hands-on

experience with many of the tools needed to build data products,
and have a better understanding of how to build scalable machine
learning pipelines in a cloud environment.
x Preface
0.3 Code Examples

Since the focus of this book is to get readers hands on with Python
code, I have provided code listings for a subset of the chapters on
GitHub. The following URL provides listings for code examples
that work well in a Jupyter environment:
• https://github.com/bgweber/DS_Production
Due to formatting restrictions, many of the code snippets in this

book break commands into multiple lines while omitting the con-
tinuation operator (\). To get code blocks to work in Jupyter or an-
other Python coding environment, you may need to remove these
line breaks. The code samples in the notebooks listed above do
not add these line breaks and can be executed without modifi-
cation, excluding credential and IP changes. This book uses the
terms scikit-learn and sklearn interchangeably with sklearn used
explicitly in Section 3.3.3.
0.4 Acknowledgements
I was able to author this book using Yihui Xie’s excellent book-
down package (Xie, 2015). For the design, I used Shashi Kumar’s
template2 available under the Creative Commons 4.0 license. The
book cover uses Cédric Franchetti’s image from pxhere3 .
This book was last updated on December 31, 2019.
2
https://bit.ly/2MjFDgV
3
https://pxhere.com/en/photo/1417846
1
Introduction
Putting predictive models into production is one of the most di-

rect ways that data scientists can add value to an organization. By
learning how to build and deploy scalable model pipelines, data sci-
entists can own more of the model production process and rapidly
deliver data products. Building data products is more than just
putting code into production, it also includes DevOps and lifecy-
cle management of live systems.
Throughout this book, we’ll cover different cloud environments
and tools for building scalable data and model pipelines. The goal
is to provide readers with the opportunity to get hands on and
start building experience with a number of different tools. While
this book is targeted at analytics practitioners with prior Python
experience, we’ll walk through examples from start to finish, but
won’t dig into the details of the programming language itself.
The role of data science is constantly transforming and adding
new specializations. Data scientists that build production-grade
services are often called applied scientists. Their goal is to build
systems that are scalable and robust. In order to be scalable, we
need to use tools that can parallelize and distribute code. Paralleliz-
ing code means that we can perform multiple tasks simultaneously,
and distributing code means that we can scale up the number of
machines needed in order to accomplish a task. Robust services
are systems that are resilient and can recover from failure. While
the focus of this book is on scalability rather than robustness, we
will cover monitoring systems in production and discuss measuring
model performance over time.
During my career as a data scientist, I’ve worked at a number of
video game companies and have had experience putting propensity
1
2 1 Introduction
models, lifetime-value predictions, and recommendation systems

into production. Overall, this process has become more stream-
lined with the development of tools such as PySpark, which enable
data scientists to more rapidly build end-to-end products. While
many companies now have engineering teams with machine learn-
ing focuses, it’s valuable for data scientists to have broad expertise
in productizing models. Owning more of the process means that a
data science team can deliver products quicker and iterate much
more rapidly.
Data products are useful for organizations, because they can pro-
vide personalization for the user base. For example, the recommen-
dation system that I designed for EverQuest Landmark1 provided
curated content for players from a marketplace with thousands
of user-created items. The goal of any data product should be
creating value for an organization. The recommendation system
accomplished this goal by increasing the revenue generated from
user-created content. Propensity models, which predict the likeli-
hood of a user to perform an action, can also have a direct impact
on core metrics for an organization, by enabling personalized ex-
periences that increase user engagement.
The process used to productize models is usually unique for each
organization, because of different cloud environments, databases,
and product organizations. However, many of the same tools are
used within these workflows, such as SQL and PySpark. Your or-
ganization may not be using the same data ecosystem as these
examples, but the methods should transfer to your use cases.
In this chapter, we will introduce the role of applied science and
motivate the usage of Python for building data products, discuss
different cloud and coding environments for scaling up data science,
introduce the data sets and types of models used throughout the
book, and introduce automated feature engineering as a step to
include in data science workflows.
1
https://bit.ly/2YFlYPg
1.1 Applied Data Science 3
1.1 Applied Data Science

Data science is a broad discipline with many different specializa-
tions. One distinction that is becoming common is product data
science and applied data science. Product data scientists are typ-
ically embedded on a product team, such as a game studio, and
provide analysis and modeling that helps the team directly im-
prove the product. For example, a product data scientist might
find an issue with the first-time user experience in a game, and
make recommendations such as which languages to focus on for
localization to improve new user retention.
Applied science is at the intersection of machine learning engineer-
ing and data science. Applied data scientists focus on building data
products that product teams can integrate. For example, an ap-
plied scientist at a game publisher might build a recommendation
service that different game teams can integrate into their products.
Typically, this role is part of a central team that is responsible for
owning a data product. A data product is a production system
that provides predictive models, such as identifying which items a
player is likely to buy.
Applied scientist is a job title that is growing in usage across tech
companies including Amazon, Facebook, and Microsoft. The need
for this type of role is growing, because a single applied scientist
can provide tremendous value to an organization. For example, in-
stead of having product data scientists build bespoke propensity
models for individual games, an applied scientist can build a scal-
able approach that provides a similar service across a portfolio of
games. At Zynga, one of the data products that the applied data
science team built was a system called AutoModel2 , which pro-
vided several propensity models for all of our games, such as the
likelihood for a specific player to churn.
There’s been a few developments in technology that have made
applied science a reality. Tools for automated feature engineering,
2
https://ubm.io/2KdYRDq
4 1 Introduction
such as deep learning, and scalable computing environments, such

as PySpark, have enabled companies to build large scale data prod-
ucts with smaller team sizes. Instead of hiring engineers for data
ingestion and warehousing, data scientists for predictive modeling,
and additional engineers for building a machine learning infras-
tructure, you can now use managed services in the cloud to enable
applied scientists to take on more of the responsibilities previously
designated to engineering teams.
One of the goals of this book is to help data scientists make the
transition to applied science, by providing hands-on experience
with different tools that can be used for scalable compute and
standing up services for predictive models. We will work through
different tools and cloud environments to build proof of concepts
for data products that can translate to production environments.
1.2 Python for Scalable Compute

Python is quickly becoming the de facto language for data science.
In addition to the huge library of packages that provide useful
functionality, one of the reasons that the language is becoming
so popular is that it can be used for building scalable data and
predictive model pipelines. You can use Python on your local ma-
chine and build predictive models with scikit-learn, or use environ-
ments such as Dataflow and PySpark to build distributed systems.
While these different environments use different libraries and pro-
gramming paradigms, it’s all in the same language of Python. It’s
no longer necessary to translate an R script into a production
language such as Java, you can use the same language for both
development and production of predictive models.
It took me awhile to adopt Python as my data science language
of choice. Java had been my preferred language, regardless of task,
since early in my undergraduate career. For data science tasks, I
used tools like Weka to train predictive models. I still find Java to
be useful when building data pipelines, and it’s great to know in
1.2 Python for Scalable Compute 5
order to directly collaborate with engineering teams on projects. I

later switched to R while working at Electronic Arts, and found
the transition to an interactive coding environment to be quite
useful for data science. One of the features I really enjoyed in R is
R Markdown, which you can use to write documents with inline
code. In fact, this entire book was written using an extension of R
Markdown called bookdown (Xie, 2019). I later switched to using
R within Jupyter notebooks and even wrote a book on using R
and Java for data science at a startup (Weber, 2018).
When I started working at Zynga in 2018, I adopted Python and
haven’t looked back. It took a bit of time to get used to the new
language, but there are a number of reasons that I wanted to learn
Python:
• Momentum: Many teams are already using Python for produc-
tion, or portions of their data pipelines. It makes sense to also
use Python for performing analysis tasks.
• PySpark: R and Java don’t provide a good transition to author-
ing Spark tasks interactively. You can use Java for Spark, but
it’s not a good fit for exploratory work, and the transition from
Python to PySpark seems to be the most approachable way to
learn Spark.
• Deep Learning: I’m interested in deep learning, and while there
are R bindings for libraries such as Keras, it’s better to code in
the native language of these libraries. I previously used R to
author custom loss functions, and debugging errors was prob-
lematic.
• Libraries: In addition to the deep learning libraries offered for
Python, there’s a number of other useful tools including Flask
and Bokeh. There’s also notebook environments that can scale
including Google’s Colaboratory and AWS SageMaker.
To ease the transition from R to Python, I took the following steps:
• Focus on outcomes, not semantics: Instead of learning
about all of the fundamentals of the language, I first focused
on doing in Python what I already knew how to do in other
languages, such as training a logistic regression model.
6 1 Introduction
• Learn the ecosystem, not the language: I didn’t limit myself

to the base language when learning, and instead jumped right in
to using Pandas and scikit-learn.
• Use cross-language libraries: I was already familiar working
with Keras and Plotly in R, and used knowledge of these libraries
to bootstrap learning Python.
• Work with real-world data: I used the data sets provided by
Google’s BigQuery to test out my scripts on large-scale data.
• Start locally if possible: While one of my goals was to learn
PySpark, I first focused on getting things up and running on my
local machine before moving to cloud ecosystems.
There are many situations where Python is not the best choice for a
specific task, but it does have broad applicability when prototyping
models and building scalable model pipelines. Because of Python’s
rich ecosystem, we will be using it for all of the examples in this
book.
1.3 Cloud Environments

In order to build scalable data science pipelines, it’s necessary to
move beyond single machine scripts and move to clusters of ma-
chines. While this is possible to do with an on-premise setup, a
common trend is using cloud computing environments in order to
achieve large-scale processing. There’s a number of different op-
tions available, with the top three platforms currently being Ama-
zon Web Services (AWS), Google Cloud Platform (GCP), and Mi-
crosoft Azure.
Most cloud platforms offer free credits for getting started. GCP
offers a $300 credit for new users to get hands on with their tools,
while AWS provides free-tier access for many services. In this book,
we’ll get hands on with both AWS and GCP, with little to no cost
involved.
1.3 Cloud Environments 7
1.3.1 Amazon Web Services (AWS)

AWS is currently the largest cloud provider, and this dominance
has been demonstrated by the number of gaming companies using
the platform. I’ve had experience working with AWS at Electronic
Arts, Twitch, and Zynga. This platform has a wide range of tools
available, but getting these components to work well together gen-
erally takes additional engineering work versus GCP.
With AWS, you can use both self-hosted and managed solutions
for building data pipelines. For example, the managed option for
messaging on AWS is Kinesis, while the self-hosted option is Kafka.
We’ll walk through examples with both of these options in Chapter
8. There’s typically a tradeoff between cost and DevOps when
choosing between self-hosted and managed options.
The default database to use on AWS is Redshift, which is a colum-
nar database. This option works well as a data warehouse, but it
doesn’t scale well to data lake volumes. It’s common for organiza-
tions to set up a separate data lake and data warehouse on AWS.
For example, it’s possible to store data on S3 and use tools such
as Athena to provide data lake functionality, while using Redshift
as a solution for a data warehouse. This approach has worked well
in the past, but it creates issues when building large-scale data
science pipelines. Moving data in and out of a relational database
can be a bottleneck for these types of workflows. One of the so-
lutions to this bottleneck is to use vendor solutions that separate
storage from compute, such as Snowflake or Delta Lake.
The first component we’ll work with in AWS is Elastic Com-
pute (EC2) instances. These are individual virtual machines that
you can spin up and provision for any necessary task. In section
1.4.1, we’ll show how to set up an instance that provides a re-
mote Jupyter environment. EC2 instances are great for getting
started with tools such as Flask and Gunicorn, and getting started
with Docker. To scale up beyond individual instances, we’ll explore
Lambda functions and Elastic Container Services.
To build scalable pipelines on AWS, we’ll focus on PySpark as the
main environment. PySpark enables Python code to be distributed
8 1 Introduction
across a cluster of machines, and vendors such as Databricks pro-

vide managed environments for Spark. Another option that is avail-
able only on AWS is SageMaker, which provides a Jupyter note-
book environment for training and deploying models. We are not
covering SageMaker in this book, because it is specific to AWS
and currently supports only a subset of predictive models. Instead,
we’ll explore tools such as MLflow.
1.3.2 Google Cloud Platform (GCP)

GCP is currently the third largest cloud platform provider, and
offers a wide range of managed tools. It’s currently being used by
large media companies such as Spotify, and within the games in-
dustry being used by King and Niantic. One of the main benefits of
using GCP is that many of the components can be wired together
using Dataflow, which is a tool for building batch and streaming
data pipelines. We’ll create a batch model pipeline with Dataflow
in Chapter 7 and a streaming pipeline in Chapter 8.
Google Cloud Platform currently offers a smaller set of tools than
AWS, but there is feature parity for many common tools, such as
PubSub in place of Kinesis, and Cloud Functions in place of AWS
Lambda. One area where GCP provides an advantage is BigQuery
as a database solution. BigQuery separates storage from compute,
and can scale to both data lake and data warehouse use cases.
Dataflow is one of the most powerful tools for data scientists that
GCP provides, because it empowers a single data scientist to build
large-scale pipelines with much less effort than other platforms.
It enables building streaming pipelines that connect PubSub for
messaging, BigQuery for analytics data stores, and BigTable for
application databases. It’s also a managed solution that can au-
toscale to meet demand. While the original version of Dataflow
was specific to GCP, it’s now based on the Apache Beam library
which is portable to other platforms.
1.4 Coding Environments 9
1.4 Coding Environments

There’s a variety of options for writing Python code in order to do
data science. The best environment to use likely varies based on
what you are building, but notebook environments are becoming
more and more common as the place to write Python scripts. The
three types of coding environments I’ve worked with for Python
are IDEs, text editors, and notebooks.
If you’re used to working with an IDE, tools like PyCharm and
Rodeo are useful editors and provide additional tools for debugging
versus other options. It’s also possible to write code in text editors
such as Sublime and then run scripts via the command line. I
find this works well for building web applications with Flask and
Dash, where you need to have a long running script that persists
beyond the scope of running a cell in a notebook. I now perform
the majority of my data science work in notebook environments,
and this covers exploratory analysis and productizing models.
I like to work in coding environments that make it trivial to share
code and collaborate on projects. Databricks and Google Colab
are two coding environments that provide truly collaborative note-
books, where multiple data scientists can simultaneously work on
a script. When using Jupyter notebooks, this level of real-time
collaboration is not currently supported, but it’s good practice to
share notebooks in version control systems such as GitHub for
sharing work.
In this book, we’ll use only the text editor and notebook environ-
ments for coding. For learning how to build scalable pipelines, I
recommend working on a remote machine, such as EC2, to become
more familiar with cloud environments, and to build experience
setting up Python environments outside of your local machine.
1.4.1 Jupyter on EC2

To get experience with setting up a remote machine, we’ll start by
setting up a Jupyter notebook environment on a EC2 instance in
10 1 Introduction
FIGURE 1.1: Public and Private IPs on EC2.
AWS. The result is a remote machine that we can use for Python
scripting. Accomplishing this task requires spinning up an EC2
instance, configuring firewall settings for the EC2 instance, con-
necting to the instance using SSH, and running a few commands
to deploy a Jupyter environment on the machine.
The first step is to set up an AWS account and log into the AWS
management console. AWS provides a free account option with
free-tier access to a number of services including EC2. Next, pro-
vision a machine using the following steps:
1. Under “Find Services”, search for EC2

2. Click “Launch Instance”
3. Select a free-tier Amazon Linux AMI
4. Click “Review and Launch”, and then “Launch”
5. Create a key pair and save to your local machine
6. Click “Launch Instances”
7. Click “View Instances”
The machine may take a few minutes to provision. Once the ma-
chine is ready, the instance state will be set to “running”. We can
now connect to the machine via SSH. One note on the different
AMI options is that some of the configurations are set up with
Python already installed. However, this book focuses on Python 3
and the included version is often 2.7.
There’s two different IPs that you need in order to connect to the
machine via SSH and later connect to the machine via web browser.
The public and private IPs are listed under the “Description” tab
1.4 Coding Environments 11
FIGURE 1.2: SSH connection settings.
as shown in Figure 1.1. To connect to the machine via SSH we’ll

use the Public IP (54.87.230.152). For connecting to the machine,
you’ll need to use an SSH client such as Putty if working in a
Windows environment. For Linux and Mac OS, you can use ssh
via the command line. To connect to the machine, use the user
name “ec2-user” and the key pair generated when launching the
instance. An example of connecting to EC2 using the Bitvise client
on Windows is shown in Figure 1.2.
Once you connect to the machine, you can check the Python ver-
sion by running python --version. On my machine, the result was
2.7.16, meaning that additional setup is needed in order to up-
grade to Python 3. We can run the following commands to install
Python 3, pip, and Jupyter.
sudo yum install -y python37

python3 --version
curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
sudo python3 get-pip.py
12 1 Introduction
pip --version
pip install --user jupyter
The two version commands are to confirm that the machine is

pointing at Python 3 for both Python and pip. Once Jupyter is
installed, we’ll need to set up a firewall restriction so that we can
connect directly to the machine on port 8888, where Jupyter runs
by default. This approach is the quickest way to get connected to
the machine, but it’s advised to use SSH tunneling to connect to
the machine rather than a direct connection over the open web.
You can open up port 8888 for your local machine by performing
the following steps from the EC2 console:
1. Select your EC2 instance

2. Under “Description”, select security groups
3. Click “Actions” -> “Edit Inbound Rules”
4. Add a new rule: change the port to 8888, select “My IP”
5. Click “Save”
We can now run and connect to Jupyter on the EC2 machine. To

launch Jupyer, run the command shown below while replacing the
IP with your EC2 instance’s Private IP. It is necessary to specify
the --ip parameter in order to enable remote connections to Jupyer,
as incoming traffic will be routed via the private IP.
jupyter notebook --ip 172.31.53.82
When you run the jupyter notebook command, you’ll get a URL
with a token that can be used to connect to the machine. Be-
fore entering the URL into your browser, you’ll need to swap the
Private IP output to the console with the Public IP of the EC2
instance, as shown in the snippet below.
# Original URL
The Jupyter Notebook is running at:
1.5 Datasets 13
FIGURE 1.3: Jupyter Notebook on EC2.
http://172.31.53.82:8888/?token=
98175f620fd68660d26fa7970509c6c49ec2afc280956a26
# Swap Private IP with Public IP

http://54.87.230.152:8888/?token=
98175f620fd68660d26fa7970509c6c49ec2afc280956a26
You can now paste the updated URL into your browser to connect
to Jupyter on the EC2 machine. The result should be a Jupyer
notebook fresh install with a single file get-pip.py in the base di-
rectory, as shown in Figure 1.3. Now that we have a machine set
up with Python 3 and Jupyter notebook, we can start exploring
different data sets for data science.
1.5 Datasets
To build scalable data pipelines, we’ll need to switch from using
local files, such as CSVs, to distributed data sources, such as Par-
quet files on S3. While the tools used across cloud platforms to
load data vary significantly, the end result is usually the same,
which is a dataframe. In a single machine environment, we can use
Pandas to load the dataframe, while distributed environments use
different implementations such as Spark dataframes in PySpark.
This section will introduce the data sets that we’ll explore through-
out the rest of the book. In this chapter we’ll focus on loading the
14 1 Introduction
data using a single machine, while later chapters will present dis-
tributed approaches. While most of the data sets presented here
can be downloaded as CSV files as read into Pandas using read_csv,
it’s good practice to develop automated workflows to connect to di-
verse data sources. We’ll explore the following datasets throughout
this book:
• Boston Housing: Records of sale prices of homes in the Boston
housing market back in 1980.
• Game Purchases: A synthetic data set representing games pur-
chased by different users on XBox One.
• Natality: One of BigQuery’s open data sets on birth statistics
in the US over multiple decades.
• Kaggle NHL: Play-by-play events from professional hockey
games and game statistics over the past decade.
The first two data sets are single commands to load, as long as
you have the required libraries installed. The Natality and Kaggle
NHL data sets require setting up authentication files before you
can programmatically pull the data sources into Pandas.
The first approach we’ll use to load a data set is to retrieve it di-
rectly from a library. Multiple libraries include the Boston housing
data set, because it is a small data set that is useful for testing out
regression models. We’ll load it from scikit-learn by first running
pip from the command line:
pip install --user pandas

pip install --user sklearn
Once scikit-learn is installed, we can switch back to the Jupyter

notebook to explore the data set. The code snippet below shows
how to load the scikit-learn and Pandas libraries, load the Boston
data set as a Pandas dataframe, and display the first 5 records.
The result of running these commands is shown in Figure 1.4.
from sklearn.datasets import load_boston

import pandas as pd
1.5 Datasets 15
FIGURE 1.4: Boston Housing data set.
data, target = load_boston(True)

bostonDF = pd.DataFrame(data, columns=load_boston().feature_names)
bostonDF['label'] = target
bostonDF.head()
The second approach we’ll use to load a data set is to fetch it

from the web. The CSV for the Games data set is available as a
single file on GitHub. We can fetch it into a Pandas dataframe by
using the read_csv function and passing the URL of the file as a
parameter. The result of reading the data set and printing out the
first few records is shown in Figure 1.5.
gamesDF = pd.read_csv("https://github.com/bgweber/
Twitch/raw/master/Recommendations/games-expand.csv")
gamesDF.head()
Both of these approaches are similar to downloading CSV files and

reading them from a local directory, but by using these methods
we can avoid the manual step of downloading files.
1.5.1 BigQuery to Pandas

One of the ways to automate workflows authored in Python is to
directly connect to data sources. For databases, you can use con-
nectors based on JDBC or native connectors, such as the bigquery
module provided by the Google Cloud library. This connector en-
ables Python applications to send queries to BigQuery and load
the results as a Pandas dataframe. This process involves setting up
16 1 Introduction
FIGURE 1.5: Game Purchases data set.
a GCP project, installing the prerequisite Python libraries, setting

up the Google Cloud command line tools, creating GCP creden-
tials, and finally sending queries to BigQuery programmatically.
If you do not already have a GCP account set up, you’ll need to
create a new account3 . Google provides a $300 credit for getting
up and running with the platform. The first step is to install the
Google Cloud library by running the following steps:
pip install --user google-cloud-bigquery

pip install --user matplotlib
Next, we’ll need to set up the Google Cloud command line tools,
in order to set up credentials for connecting to BigQuery. While
the files to use will vary based on the current release4 , here are the
steps I ran on the command line:
curl -O https://dl.google.com/dl/cloudsdk/channels/
rapid/downloads/google-cloud-sdk-255.0.0-
linux-x86_64.tar.gz
tar zxvf google-cloud-sdk-255.0.0-linux-x86_64.tar.gz
google-cloud-sdk
./google-cloud-sdk/install.sh
3
https://cloud.google.com/gcp
4
https://cloud.google.com/sdk/install
1.5 Datasets 17
Once the Google Cloud command line tools are installed, we can
set up credentials for connecting to BigQuery:
gcloud config set project project_name

gcloud auth login
gcloud init
gcloud iam service-accounts create dsdemo
gcloud projects add-iam-policy-binding your_project_id
--member "serviceAccount:dsdemo@your_project_id.iam.
gserviceaccount.com" --role "roles/owner"
gcloud iam service-accounts keys
create dsdemo.json --iam-account
dsdemo@your_project_id.iam.gserviceaccount.com
export GOOGLE_APPLICATION_CREDENTIALS=
/home/ec2-user/dsdemo.json
You’ll need to substitute project_name with your project name,

your_project_id with your project ID, and dsdemo with your desired
service account name. The result is a json file with credentials
for the service account. The export command at the end of this
process tells Google Cloud where to find the credentials file.
Setting up credentials for Google Cloud is involved, but generally
only needs to be performed once. Now that credentials are config-
ured, it’s possible to directly query BigQuery from a Python script.
The snippet below shows how to load a BigQuery client, send a
SQL query to retrieve 10 rows from the natality data set, and pull
the results into a Pandas dataframe. The resulting dataframe is
shown in Figure 1.6.
from google.cloud import bigquery

client = bigquery.Client()
sql = """
SELECT *
FROM `bigquery-public-data.samples.natality`
limit 10
"""
18 1 Introduction
FIGURE 1.6: Previewing the BigQuery data set.
natalityDF = client.query(sql).to_dataframe()
natalityDF.head()
1.5.2 Kaggle to Pandas

Kaggle is a data science website that provides thousands of open
data sets to explore. While it is not possible to pull Kaggle data
sets directly into Pandas dataframes, we can use the Kaggle library
to programmatically download CSV files as part of an automated
workflow.
The quickest way to get set up with this approach is to create an
account on Kaggle5 . Next, go to the account tab of your profile
and select ‘Create API Token’, download and open the file, and
then run vi .kaggle/kaggle.json on your EC2 instance to copy over
the contents to your remote machine. The result is a credential file
you can use to programmatically download data sets. We’ll explore
the NHL (Hockey) data set by running the following commands:
pip install kaggle --user
kaggle datasets download martinellis/nhl-game-data

unzip nhl-game-data.zip
chmod 0600 *.csv
5
https://www.kaggle.com
1.6 Prototype Models 19
FIGURE 1.7: NHL Kaggle data set.
These commands will download the data set, unzip the files into
the current directory, and enable read access on the files. Now
that the files are downloaded on the EC2 instance, we can load
and display the Game data set, as shown in Figure 1.7. This data
set includes different files, where the game file provides game-level
summaries and the game_plays file provides play-by-play details.
import pandas as pd
nhlDF = pd.read_csv('game.csv')
nhlDF.head()
We walked through a few different methods for loading data sets

into a Pandas dataframe. The common theme with these different
approaches is that we want to avoid manual steps in our workflows,
in order to automate pipelines.
1.6 Prototype Models

Machine learning is one of the most important steps in the pipeline
of a data product. We can use predictive models to identify which
users are most likely to purchase an item, or which users are most
likely to stop using a product. The goal of this section is to present
simple versions of predictive models that we’ll later scale up in
more complex pipelines. This book will not focus on state-of-the-
20 1 Introduction
art models, but instead cover tools that can be applied to a variety
of different machine learning algorithms.
The library to use for implementing different models will vary
based on the cloud platform and execution environment being used
to deploy a model. The regression models presented in this section
are built with scikit-learn, while the models we’ll build out with
PySpark use MLlib.
1.6.1 Linear Regression

Regression is a common task for supervised learning, such as pre-
dicting the value of a home, and linear regression is a useful algo-
rithm for making predictions for these types of problems. scikit-
learn provides both linear and logistic regression models for mak-
ing predictions. We’ll start by using the LinearRegression class in
scikit-learn to predict home prices for the Boston housing data set.
The code snippet below shows how to split the Boston data set into
different training and testing data sets and separate data (train_x)
and label (train_y) objects, create and fit a linear regression model,
and calculate error metrics on the test data set.
from sklearn.linear_model import LinearRegression

from sklearn.model_selection import train_test_split
# See Section 1.4 (Boston Hosing data set)

bostonDF = ...
x_train, x_test, y_train, y_test = train_test_split(

bostonDF.drop(['label'],axis=1),bostonDF['label'],test_size=0.3)
model = LinearRegression()
model.fit(x_train, y_train)
print("R^2: " + str(model.score(x_test, y_test)))

print("Mean Error: " + str(sum(
abs(y_test - model.predict(x_test) ))/y_test.count()))
The train_test_split function is used to split up the data set into

70% train and 30% holdout data sets. The first parameter is the
data attributes from the Boston dataframe, with the label dropped,
and the second parameter is the labels from the dataframe. The
two commands at the end of the script calculate the R-squared
value based on Pearson correlation, and the mean error is defined
as the mean difference between predicted and actual home prices.
The output of this script was an R2 value of 0.699 and mean error of
3.36. Since house prices in this data set are divided by a thousand,
the mean error is $3.36k.
We now have a simple model that we can productize in a number of
different environments. In later sections and chapters, we’ll explore
methods for scaling features, supporting more complex regression
models, and automating feature generation.
1.6.2 Logistic Regression

Logistic regression is a supervised classification algorithm that is
useful for predicting which users are likely to perform an action,
such as purchasing a product. Using scikit-learn, the process is
similar to fitting a linear regression model. The main differences
from the prior script are the data set being used, the model object
instantiated (LogisticRegression), and using the predict_proba func-
tion to calculate error metrics. This function predicts a probability
in the continuous range of [0,1] rather than a specific label. The
snippet below predicts which users are likely to purchase a specific
game based on prior games already purchased:
from sklearn.linear_model import LogisticRegression

from sklearn.metrics import roc_auc_score
import pandas as pd
# Games data set

gamesDF = pd.read_csv("https://github.com/bgweber/Twitch/raw/
master/Recommendations/games-expand.csv")
22 1 Introduction

gamesDF.drop(['label'],axis=1),gamesDF['label'],test_size=0.3)
model = LogisticRegression()
print("Accuracy: " + str(model.score(x_test, y_test)))

print("ROC: " + str(roc_auc_score(y_test,
model.predict_proba(x_test)[:, 1] )))
The output of this script is two metrics that describe the perfor-
mance of the model on the holdout data set. The accuracy metric
describes the number of correct predictions over the total num-
ber of predictions, and the ROC metric describes the number of
correctly classified outcomes based on different model thresholds.
ROC is a useful metric to use when the different classes being pre-
dicted are imbalanced, with noticeably different sizes. Since most
players are unlikely to buy a specific game, ROC is a good metric
to utilize for this use case. When I ran this script, the result was
an accuracy of 86.6% and an ROC score of 0.757.
Linear and logistic regression models with scikit-learn are a good
starting point for many machine learning projects. We’ll explore
more complex models in this book, but one of the general strategies
I take as a data scientist is to quickly deliver a proof of concept,
and then iterate and improve a model once it is shown to provide
value to an organization.
1.6.3 Keras Regression

While I generally recommend starting with simple approaches
when building model pipelines, deep learning is becoming a pop-
ular tool for data scientists to apply to new problems. It’s great
to explore this capability when tackling new problems, but scaling
up deep learning in data science pipelines presents a new set of
challenges. For example, PySpark does not currently have a na-
tive way of distributing the model application phase to big data.
There’s plenty of books for getting started with deep learning in

Python, such as (Chollet, 2017).
In this section, we’ll repeat the same task from the prior section,
which is predicting which users are likely to buy a game based on
their prior purchases. Instead of using a shallow learning approach
to predict propensity scores, we’ll use the Keras framework to build
a neural network for predicting this outcome. Keras is a general
framework for working with deep learning implementations. We
can install these dependencies from the command line:
pip install --user tensorflow==1.14.0

pip install --user keras==2.2.4
This process can take awhile to complete, and based on your en-
vironment may run into installation issues. It’s recommended to
verify that the installation worked by checking your Keras version
in a Jupyter notebook:
import tensorflow as tf
import keras
from keras import models, layers
import matplotlib.pyplot as plt
keras.__version__
The general process for building models with Keras is to set up

the structure of the model, compile the model, fit the model, and
evaluate the model. We’ll start with a simple model to provide a
baseline, which is shown in the snippet below. This code creates a
network with an input layer, a dropout layer, a hidden layer, and
an output layer. The input to the model is 10 binary variables that
describe prior games purchased, and the output is a prediction of
the likelihood to purchase a specified game.

gamesDF.drop(['label'], axis=1),gamesDF['label'],test_size=0.3)
24 1 Introduction
# define the network structure

model = models.Sequential()
model.add(layers.Dense(64, activation='relu', input_shape=(10,)))
model.add(layers.Dropout(0.1))
model.add(layers.Dense(64, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))
# define ROC AUC as a metric

def auc(y_true, y_pred):
auc = tf.metrics.auc(y_true, y_pred)[1]
keras.backend.get_session().run(
tf.local_variables_initializer())
return auc
# compile and fit the model

model.compile(optimizer='rmsprop',
loss='binary_crossentropy', metrics=[auc])
history = model.fit(x_train, y_train, epochs=100, batch_size=100,
validation_split = .2, verbose=0)
Since the goal is to identify the likelihood of a player to purchase a

game, ROC is a good metric to use to evaluate the performance of
a model. Keras does not support this directly, but we can define a
custom metrics function that wraps the auc functionality provided
by TensorFlow.
Next, we specify how to optimize the model. We’ll use rmsprop for
the optimizer and binary_crossentropy for the loss function.The last
step is to train the model. The code snippet shows how to fit the
model using the training data set, 100 training epochs with a batch
size of 100, and a cross validation split of 20%. This process can
take awhile to run if you increase the number of epochs or decrease
the batch size. The validation set is sampled from only the training
data set.
The result of this process is a history object that tracks the loss
and metrics on the training and validation data sets. The code
FIGURE 1.8: ROC AUC metrics for the training process.
snippet below shows how to plot these values using Matplotlib.

The output of this step is shown in Figure 1.8. The plot shows that
both the training and validation data sets leveled off at around a
0.82 AUC metric during the model training process. To compare
this approach with the logistic regression results, we’ll evaluate the
performance of the model on the holdout data set.
loss = history.history['auc']
val_loss = history.history['val_auc']
epochs = range(1, len(loss) + 1)
plt.figure(figsize=(10,6) )
plt.plot(epochs, loss, 'bo', label='Training AUC')
plt.plot(epochs, val_loss, 'b', label='Validation AUC')
plt.legend()
plt.show()
To measure the performance of the model on the test data set, we

can use the evaluate function to measure the ROC metric. The code
snippet below shows how to perform this task on our training data
set, which results in an ROC AUC value of 0.816. This is noticeably
better than the performance of the logistic regression model, with
an AUC value of 0.757, but using other shallow learning methods
26 1 Introduction
such as random forests or XGBoost would likely perform much

better on this task.
results = model.evaluate(x_test, y_test, verbose = 0)

print("ROC: " + str(results[1]))
1.7 Automated Feature Engineering

Automated feature engineering is a powerful tool for reducing the
amount of manual work needed in order to build predictive mod-
els. Instead of a data scientist spending days or weeks coming up
with the best features to describe a data set, we can use tools
that approximate this process. One library I’ve been working with
to implement this step is FeatureTools. It takes inspiration from
the automated feature engineering process in deep learning, but
is meant for shallow learning problems where you already have
structured data, but need to translate multiple tables into a single
record per user. The library can be installed as follows:
sudo yum install gcc

sudo yum install python3-devel
pip install --user framequery
pip install --user fsspec
pip install --user featuretools
In addition to this library, I loaded the framequery library, which

enables writing SQL queries against dataframes. Using SQL to
work with dataframes versus specific interfaces, such as Pandas, is
useful when translating between different execution environments.
The task we’ll apply the FeatureTools library to is predicting which
games in the Kaggle NHL data set are postseason games. We’ll
make this prediction based on summarizations of the play events
that are recorded for each game. Since there can be hundreds of
play events per game, we need a process for aggregating these into
1.7 Automated Feature Engineering 27
a single summary per game. Once we aggregate these events into

a single game record, we can apply a logistic regression model to
predict whether the game is regular or postseason.
The first step we’ll perform is loading the data sets and performing
some data preparation, as shown below. After loading the data sets
as Pandas dataframes, we drop a few attributes from the plays
object, and fill any missing attributes with 0.
import pandas as pd
game_df = pd.read_csv("game.csv")
plays_df = pd.read_csv("game_plays.csv")
plays_df = plays_df.drop(['secondaryType', 'periodType',

'dateTime', 'rink_side'], axis=1).fillna(0)
To translate the play events into a game summary, we’ll first 1-

hot econde two of the attributes in the plays dataframe, and then
perform deep feature synthesis. The code snippet below shows how
to perform the first step, and uses FeatureTools to accomplish this
task. The result is quite similar to using the get_dummies function
in Pandas, but this approach requires some additional steps.
The base representation in FeatureTools is an EntitySet, which
describes a set of tables and the relationships between them,
which is similar to defining foreign key constraints. To use the
encode_features function, we need to first translate the plays
dataframe into an entity. We can create an EntitySet directly from
the plays_df object, but we also need to specify which attributes
should be handled as categorical, using the variable_types dictio-
nary parameter.
import featuretools as ft
from featuretools import Feature
es = ft.EntitySet(id="plays")
28 1 Introduction
FIGURE 1.9: The 1-hot encoded Plays dataframe.
es = es.entity_from_dataframe(entity_id="plays",dataframe=plays_df
,index="play_id", variable_types = {
"event": ft.variable_types.Categorical,
"description": ft.variable_types.Categorical })
f1 = Feature(es["plays"]["event"])
f2 = Feature(es["plays"]["description"])
encoded, defs = ft.encode_features(plays_df, [f1, f2], top_n=10)

encoded.reset_index(inplace=True)
encoded.head()
Next, we pass a list of features to the encode_features function,

which returns a new dataframe with the dummy variables and a
defs object that describes how to translate an input dataframe
into the 1-hot encoded format. For pipelines later on in this book,
where we need to apply transformations to new data sets, we’ll
store a copy of the defs object for later use. The result of applying
this transformation to the plays dataframe is shown in Figure 1.9.
The next step is to aggregate the hundreds of play events per game
into single game summaries, where the resulting dataframe has
a single row per game.To accomplish this task, we’ll recreate the
EntitySet from the prior step, but use the 1-hot encoded dataframe
as the input. Next, we use the normalize_entity function to describe
games as a parent object to plays events, where all plays with the
same game_id are grouped together. The last step is to use the dfs
function to perform deep feature synthesis. DFS applies aggregate
calculations, such as SUM and MAX, across the different features in
1.7 Automated Feature Engineering 29
FIGURE 1.10: Generated features for the NHL data set.
the child dataframe, in order to collapse hundreds of records into

a single row.
es = es.entity_from_dataframe(entity_id="plays",
dataframe=encoded, index="play_id")
es = es.normalize_entity(base_entity_id="plays",
new_entity_id="games", index="game_id")
features,transform=ft.dfs(entityset=es,
target_entity="games",max_depth=2)
features.reset_index(inplace=True)
features.head()
The result of this process is shown in Figure 1.10. The shape of

the sampled dataframe, 5 rows by 212 columns, indicates that we
have generated hundreds of features to describe each game using
deep feature synthesis. Instead of hand coding this translation, we
utilized the FeatureTools library to automate this process.
Now that we have hundreds of features for describing a game, we
can use logistic regression to make predictions about the games.
For this task, we want to predict whether a game is regular season
or postseason, where type = 'P'. The code snippet below shows how
to use the framequery library to combine the generated features
with the initially loaded games dataframe using a SQL join. We
use the type attribute to assign a label, and then return all of the
generated features and the label. The result is a dataframe that
we can pass to scikit-learn.
30 1 Introduction
import framequery as fq
# assign labels to the generated features

features = fq.execute("""
SELECT f.*
,case when g.type = 'P' then 1 else 0 end as label
FROM features f
JOIN game_df g
on f.game_id = g.game_id
""")
We can re-use the logistic regression code from above to build a

model that predicts whether an NHL game is a regular or postsea-
son game. The updated code snippet to build a logistic regression
model with scikit-learn is shown below. We drop the game_id col-
umn before fitting the model to avoid training the model on this
attribute, which typically results in overfitting.

from sklearn.metrics import roc_auc_score
# create inputs for sklearn

y = features['label']
X = features.drop(['label', 'game_id'], axis=1).fillna(0)
# train a classifier
lr = LogisticRegression()
model = lr.fit(X, y)
# Results
print("Accuracy: " + str(model.score(X, y)))
print("ROC" + str(roc_auc_score(y,model.predict_proba(X)[:,1])))
The result of this model was an accuracy of 94.7% and an ROC

measure of 0.923. While we likely could have created a better
performing model by manually specifying how to aggregate play
1.8 Conclusion 31
events into a game summary, we were able to build a model with

good accuracy while automating much of this process.
1.8 Conclusion
Building data products is becoming an essential competency for ap-
plied data scientists. The Python ecosystem provides useful tools
for taking prototype models and scaling them up to production-
quality systems. In this chapter, we laid the groundwork for the
rest of this book by introducing the data sets, coding tools, cloud
environments, and predictive models that we’ll use to build scal-
able model pipelines. We also explored a recent Python library
called FeatureTools, which enables automating much of the fea-
ture engineering steps in a model pipeline.
In our current setup, we built a simple batch model on a single
machine in the cloud. In the next chapter, we’ll explore how to
share our models with the world, by exposing them as endpoints
on the web.
2
Models as Web Endpoints
In order for a machine learning model to be useful, you need a

way of sharing the results with other services and applications
within your organization. While you can precompute results and
save them to a database using a batch pipeline approach, it’s of-
ten necessary to respond to requests in real-time with up-to-date
information. One way of achieving this goal is by setting up a pre-
dictive model as a web endpoint that can be invoked from other
services. This chapter shows how to set up this functionality for
both scikit-learn and Keras models, and introduces Python tools
that can help scale up this functionality.
It’s good to build experience both hosting and consuming web end-
points when building out model pipelines with Python. In some
cases, a predictive model will need to pull data points from other
services before making a prediction, such as needing to pull addi-
tional attributes about a user’s history as input to feature engineer-
ing. In this chapter, we’ll focus on JSON based services, because it
is a popular data format and works well with Python’s data types.
A model as an endpoint is a system that provides a prediction in
response to a passed in set of parameters. These parameters can
be a feature vector, image, or other type of data that is used as
input to a predictive model. The endpoint then makes a prediction
and returns the results, typically as a JSON payload. The benefits
of setting up a model this way are that other systems can use the
predictive model, it provides a real-time result, and can be used
within a broader data pipeline.
In this chapter, we’ll cover calling web services using Python, set-
ting up endpoints, saving models so that they can be used in pro-
duction environments, hosting scikit-learn and Keras predictive
33
34 2 Models as Web Endpoints
models, scaling up a service with Gunicorn and Heroku, and build-

ing an interactive web application with Plotly Dash.
2.1 Web Services

Before we host a predictive model, we’ll use Python to call a web
service and to process the result. After showing how to process a
web response, we’ll set up our own service that echoes the passed
in message back to the caller. There’s a few different libraries we’ll
need to install for the examples in this chapter:
pip install --user requests

pip install --user flask
pip install --user gunicorn
pip install --user mlflow
pip install --user pillow
pip install --user dash
These libraries provide the following functionality:

• requests: Provides functions for GET and POST commands.
• flask: Enables functions to be exposed as HTTP locations.
• gunicorn: A WSGI server that enables hosting Flask apps in
production environments.
• mlflow: A model library that provides model persistence.
• pillow: A fork of the Python Imaging Library.
• dash: Enables writing interactive web apps in Python.
Many of the tools for building web services in the Python ecosys-
tem work well with Flask. For example, Gunicorn can be used to
host Flask applications at production scale, and the Dash library
builds on top of Flask.
To get started with making web requests in Python, we’ll use the
Cat Facts Heroku app1 . Heroku is a cloud platform that works
1
https://cat-fact.herokuapp.com/#/
2.1 Web Services 35
well for hosting Python applications that we’ll explore later in this
chapter. The Cat Facts service provides a simple API that provides
a JSON response containing interesting tidbits about felines. We
can use the /facts/random endpoint to retrieve a random fact using
the requests library:
import requests
result = requests.get("http://cat-fact.herokuapp.com/facts/random")
print(result)
print(result.json())
print(result.json()['text'])
This snippet loads the requests library and then uses the get func-
tion to perform an HTTP get for the passed in URL. The result
is a response object that provides a response code and payload if
available. In this case, the payload can be processed using the json
function, which returns the payload as a Python dictionary. The
three print statements show the response code, the full payload,
and the value for the text key in the returned dictionary object.
The output for a run of this script is shown below.
<Response [200]>
{'used': False, 'source': 'api', 'type': 'cat', 'deleted': False

,'_id': '591f98c5d1f17a153828aa0b', '__v': 0, 'text':
'Domestic cats purr both when inhaling and when exhaling.',
'updatedAt': '2019-05-19T20:22:45.768Z',
'createdAt': '2018-01-04T01:10:54.673Z'}
Domestic cats purr both when inhaling and when exhaling.
2.1.1 Echo Service

Before we set up a complicated environment for hosting a predic-
tive model, we’ll start with a simple example. The first service
we’ll set up is an echo application that returns the passed in mes-
sage parameter as part of the response payload. To implement this

functionality, we’ll use Flask to build a web service hosted on an
EC2 instance. This service can be called on the open web, using
the public IP of the EC2 instance. You can also run this service
on your local machine, but it won’t we accessible over the web. In
order to access the function, you’ll need to enable access on port
5000, which is covered in Section 1.4.1. The complete code for the
echo web service is shown below:
import flask
app = flask.Flask(__name__)
@app.route("/", methods=["GET","POST"])
def predict():
data = {"success": False}
# check for passed in parameters

params = flask.request.json
if params is None:
params = flask.request.args
# if parameters are found, echo the msg parameter

if "msg" in params.keys():
data["response"] = params.get("msg")
data["success"] = True
return flask.jsonify(data)
if __name__ == '__main__':
app.run(host='0.0.0.0')
The first step is loading the Flask library and creating a Flask
object using the name special variable. Next, we define a predict
function with a Flask annotation that specifies that the function
should be hosted at “/” and accessible by HTTP GET and POST
commands. The last step specifies that the application should run
using 0.0.0.0 as the host, which enables remote machines to access
2.1 Web Services 37
the application. By default, the application will run on port 5000,

but it’s possible to override this setting with the port parameter.
When running Flask directly, we need to call the run function, but
we do not want to call the command when running as a module
within another application, such as Gunicorn.
The predict function returns a JSON response based on the passed
in parameters. In Python, you can think of a JSON response as a
dictionary, because the jsonify function in Flask makes the trans-
lation between these data formats seamless. The function first de-
fines a dictionary with the success key set to False. Next, the func-
tion checks if the request.json or request.args values are set, which
indicates that the caller passed in arguments to the function, which
we’ll cover in the next code snippet. If the user has passed in a msg
parameter, the success key is set to True and a response key is set
to the msg parameter in the dictionary. The result is then returned
as a JSON payload.
Instead of running this in a Jupyter notebook, we’ll save the script
as a file called echo.py. To launch the Flask application, run python3
echo.py on the command line. The result of running this command
is shown below:
python3 echo.py
* Serving Flask app "echo" (lazy loading)
* Environment: production
WARNING: This is a development server.
Do not use it in a production deployment.
Use a production WSGI server instead.
* Debug mode: off
* Running on http://0.0.0.0:5000/ (Press CTRL+C to quit)
The output indicates that the service is running on port 5000. If

you launched the service on your local machine, you can browse
to http://localhost:5000 to call the application, and if using EC2
you’ll need to use the public IP, such as http://52.90.199.190:5000.
The result will be {"response":null,"success":false}, which indi-
cates that the service was called but that no message was provided
to the echo service.
We can pass parameters to the web service using a few different ap-
proaches. The parameters can be appended to the URL, specified
using the params object when using a GET command, or passed
in using the json parameter when using a POST command. The
snippet below shows how to perform these types of requests. For
small sets of parameters, the GET approach works fine, but for
larger parameters, such as sending images to a server, the POST
approach is preferred.
import requests
result = requests.get("http://52.90.199.190:5000/?msg=HelloWorld!")
result = requests.get("http://52.90.199.190:5000/",
params = { 'msg': 'Hello from params' })
result = requests.post("http://52.90.199.190:5000/",
json = { 'msg': 'Hello from data' })
The output of the code snippet is shown below. There are 3 JSON
responses showing that the service successfully received the mes-
sage parameter and echoed the response:
{'response': 'HelloWorld!', 'success': True}

{'response': 'Hello from params', 'success': True}
{'response': 'Hello from data', 'success': True}
In addition to passing values to a service, it can be useful to pass

larger payloads, such as images when hosting deep learning models.
One way of achieving this task is by encoding images as strings,
which will work with our existing echo service. The code snippet
2.2 Web Services 39
FIGURE 2.1: Passing an image to the echo web service.
below shows how to read in an image and perform base64 encoding

on the image before adding it to the request object. The echo
service responds with the image payload and we can use the PIL
library to render the image as a plot.

import numpy as np
from PIL import Image
import io
import base64
image = open("luna.png", "rb").read()

encoded = base64.b64encode(image)
json = {'msg': encoded})
encoded = result.json()['response']
imgData = base64.b64decode(encoded)
plt.imshow( np.array(Image.open(io.BytesIO(imgData))))
We can run the script within a Jupyter notebook. The script will
load the image and send it to the server, and then render the result
as a plot. The output of this script, which uses an image of my
in-laws’ cat, is shown in Figure 2.1. We won’t work much with
image data in this book, but I did want to cover how to use more
complex objects with web endpoints.
2.2 Model Persistence

To host a model as a web service, we need to provide a model
object for the predict function. We can train the model within
the web service application, or we can use a pre-trained model.
Model persistence is a term used for saving and loading models
within a predictive model pipeline. It’s common to train models
in a separate workflow than the pipeline used to serve the model,
such as a Flask application. In this section, we’ll save and load
both scikit-learn and Keras models, with both direct serialization
and the MLFlow library. The goal of saving and loading these
models is to make the logistic regression and deep learning models
we built in Chapter 1 available as web endpoints.
2.2.1 Scikit-Learn
We’ll start with scikit-learn, which we previously used to build a
propensity model for identifying which players were most likely to
purchase a game. A simple LogisticRegression model object can be
created using the following script:
import pandas as pd
df = pd.read_csv("https://github.com/bgweber/Twitch/
raw/master/Recommendations/games-expand.csv")
x = df.drop(['label'], axis=1)
y = df['label']
model.fit(x, y)
The default way of saving scikit-learn models is by using pickle,

which provides serialization of Python objects. You can save a
model using dump and load a model using the load function, as
2.2 Model Persistence 41
shown below. Once you have loaded a model, you can use the
prediction functions, such as predict_proba.
import pickle
pickle.dump(model, open("logit.pkl", 'wb'))
model = pickle.load(open("logit.pkl", 'rb'))

model.predict_proba(x)
Pickle is great for simple workflows, but can run into serialization
issues when your execution environment is different from your pro-
duction environment. For example, you might train models on your
local machine using Python 3.7 but need to host the models on an
EC2 instance running Python 3.6 with different library versions
installed.
MLflow is a broad project focused on improving the lifecycle of
machine learning projects. The Models component of this platform
focuses on making models deployable across a diverse range of exe-
cution environments. A key goal is to make models more portable,
so that your training environment does not need to match your
deployment environment. In the current version of MLflow, many
of the save and load functions wrap direct serialization calls, but
future versions will be focused on using generalized model formats.
We can use MLflow to save a model using sklearn.save_model and
load a model using sklearn.load_model. The script below shows how
to perform the same task as the prior code example, but uses
MLflow in place of pickle. The file is saved at the model_path lo-
cation, which is a relative path. There’s also a commented out
command, which needs to be uncommented if the code is executed
multiple times. MLflow currently throws an exception if a model
is already saved at the current location, and the rmtee command
can be used to overwrite the existing model.
import mlflow
import mlflow.sklearn
import shutil
model_path = "models/logit_games_v1"
#shutil.rmtree(model_path)
mlflow.sklearn.save_model(model, model_path)
loaded = mlflow.sklearn.load_model(model_path)
loaded.predict_proba(x)
2.2.2 Keras
Keras provides built-in functionality for saving and loading deep
learning models. We covered building a Keras model for the games
data set in Section 1.6.3. The key steps in this process are shown
in the following snippet:
import keras
# define the network structure

model.add(layers.Dense(64,activation='relu',input_shape=(10,)))

keras.backend.get_session().run(
return auc
model.compile(optimizer='rmsprop',
loss='binary_crossentropy', metrics=[auc])
2.3 Model Endpoints 43
history = model.fit(x, y, epochs=100, batch_size=100,

Once we have trained a Keras model, we can use the save and
load_model functions to persist and reload the model using the h5
file format. One additional step here is that we need to pass the
custom auc function we defined as a metric to the load function in
order to reload the model. Once the model is loaded, we can call
the prediction functions, such as evaluate.
from keras.models import load_model

model.save("games.h5")
model = load_model('games.h5', custom_objects={'auc': auc})

model.evaluate(x, y, verbose = 0)
We can also use MLflow for Keras. The save_model and load_model
functions can be used to persist Keras models. As before, we need
to provide the custom-defined auc function to load the model.
import mlflow.keras
model_path = "models/keras_games_v1"
mlflow.keras.save_model(model, model_path)
loaded = mlflow.keras.load_model(model_path,
custom_objects={'auc': auc})
loaded.evaluate(x, y, verbose = 0)
2.3 Model Endpoints

Now that we know how to set up a web service and load pre-trained
predictive models, we can set up a web service that provides a
prediction result in response to a passed-in instance. We’ll deploy

models for the games data set using scikit-learn and Keras.
2.3.1 Scikit-Learn
To use scikit-learn to host a predictive model, we’ll modify our echo
service built with Flask. The main changes to make are loading
a scikit-learn model using MLflow, parsing out the feature vector
to pass to the model from the input parameters, and adding the
model result to the response payload. The updated Flask applica-
tion for using scikit-learn is shown in the following snippet:
import pandas as pd
import mlflow
import flask
model = mlflow.sklearn.load_model(model_path)
def predict():
if "G1" in params.keys():
new_row = { "G1": params.get("G1"),"G2": params.get("G2"),
"G3": params.get("G3"),"G4": params.get("G4"),
"G9": params.get("G9"),"G10":params.get("G10")}
new_x = pd.DataFrame.from_dict(new_row,
orient = "index").transpose()
data["response"] = str(model.predict_proba(new_x)[0][1])
if __name__ == '__main__':
After loading the required libraries, we use load_model to load the

scikit-learn model object using MLflow. In this setup, the model is
loaded only once, and will not be updated unless we relaunch the
application. The main change from the echo service is creating the
feature vector that we need to pass as input to the model’s predic-
tion functions. The new_row object creates a dictionary using the
passed in parameters. To provide the Pandas row format needed
by scikit-learn, we can create a Pandas dataframe based on the dic-
tionary and then transpose the result, which creates a dataframe
with a single row. The resulting dataframe is then passed to pre-
dict_proba to make a propensity prediction for the passed in user.
The model output is added to the JSON payload under the re-
sponse key.
Similar to the echo service, we’ll need to save the app as a Python
file rather than running the code directly in Jupyter. I saved the
code as predict.py and launched the endpoint by running python3
predict.py, which runs the service on port 5000.
To test the service, we can use Python to pass in a record repre-

senting an individual user. The dictionary defines the list of games
that the user has previously purchased, and the GET command
is used to call the service. For the example below, the response
key returned a value of 0.3812. If you are running this script in a
Jupyer notebook on an EC2 instance, you’ll need to enable remote
access for the machine on port 5000. Even though the web service
and notebook are running on the same machine, we are using the
public IP to reference the model endpoint.
import requests
new_row = { "G1": 0, "G2": 0, "G3": 0, "G4": 0, "G5": 0,

"G6": 0, "G7": 0, "G8": 0, "G9": 0, "G10": 1 }
result = requests.get("http://52.90.199.190:5000/", params=new_row)

print(result.json()['response'])
2.3.2 Keras
The setup for Keras is similar to scikit-learn, but there are a few
additions that need to be made to handle the TensorFlow graph
context. We also need to redefine the auc function prior to loading
the model using MLflow. The snippet below shows the complete
code for a Flask app that serves a Keras model for the game pur-
chases data set.
The main thing to note in this script is the use of the graph object.
Because Flask uses multiple threads, we need to define the graph
used by Keras as a global object, and grab a reference to the graph
using the with statement when serving requests.
import pandas as pd
import mlflow
import mlflow.keras
import flask
import keras as k

k.backend.get_session().run(
return auc
global graph
graph = tf.get_default_graph()
model_path = "models/keras_games_v1"
model = mlflow.keras.load_model(model_path,
custom_objects={'auc': auc})
def predict():
if "G1" in params.keys():
new_row = { "G1": params.get("G1"), "G2": params.get("G2"),
"G3": params.get("G3"), "G4": params.get("G4"),
"G9": params.get("G9"), "G10": params.get("G10") }
with graph.as_default():
data["response"] = str(model.predict(new_x)[0][0])
if __name__ == '__main__':
I saved the script as keras_predict.py and then launched the Flask

app using python3 keras_predict.py. The result is a Keras model
running as a web service on port 5000. To test the script, we can
run the same script from the following section where we tested a
scikit-learn model.
2.4 Deploying a Web Endpoint

Flask is great for prototyping models as web services, but it’s not
intended to be used directly in a production environment. For
a proper deployment of a web application, you’ll want to use a
WSGI Server, which provides scaling, routing, and load balancing.
If you’re looking to host a web service that needs to handle a large
workload, then Gunicorn provides a great solution. If instead you’d
like to use a hosted solution, then Heroku provides a platform for
hosting web services written in Python. Heroku is useful for hosting
a data science portfolio, but is limited in terms of components
when building data and model pipelines.
2.4.1 Gunicorn
We can use Gunicorn to provide a WSGI server for our echo Flask
application. Using gunicorn helps separate the functionality of an
application, which we implemented in Flask, with the deployment
of an application. Gunicorn is a lightweight WSGI implementation
that works well with Flask apps.
It’s straightforward to switch form using Flask directly to using
Gunicorn to run the web service. The new command for running
the application is shown below. Note that we are passing in a bind
parameter to enable remote connections to the service.
gunicorn --bind 0.0.0.0 echo:app
The result on the command line is shown below. The main differ-
ence from before is that we now interface with the service on port
8000 rather than on port 5000. If you want to test out the service,
you’ll need to enable remote access on port 8000.
gunicorn --bind 0.0.0.0 echo:app

[INFO] Starting gunicorn 19.9.0
[INFO] Listening at: http://0.0.0.0:8000 (9509)
2.4 Deploying a Web Endpoint 49
[INFO] Using worker: sync

[INFO] Booting worker with pid: 9512
To test the service using Python, we can run the following snippet.
You’ll need to make sure that access to port 8000 is enabled, as
discussed in Section 1.4.1.
params = { 'msg': 'Hello from Gunicorn' })
The result is a JSON response with the passed in message. The

main distinction from our prior setup is that we are now using
Gunicorn, which can use multiple threads to handle load balanc-
ing, and can perform additional server configuration that is not
available when using only Flask. Configuring Gunicorn to serve
production workloads is outside the scope of this book, because
it is a hosted solution where a team needs to manage DevOps of
the system. Instead, we’ll focus on managed solutions, including
AWS Lambda and Cloud Functions in Chapter 3, where minimal
overhead is needed to keep systems operational.
2.4.2 Heroku
Now that we have a Gunicorn application, we can host it in the
cloud using Heroku. Python is one of the core languages supported
by this cloud environment. The great thing about using Heroku is
that you can host apps for free, which is great for showcasing data
science projects. The first step is to set up an account on the web
site: https://www.heroku.com/
Next, we’ll set up the command line tools for Heroku, by running
the commands shown below. There can be some complications
when setting up Heroku on an AMI EC2 instance, but downloading
and unzipping the binaries directly works around these problems.
The steps shown below download a release, extract it, and install
an additional dependency. The last step outputs the version of
Heroku installed. I got the following output: heroku/7.29.0 linux-

x64 node-v11.14.0.
wget https://cli-assets.heroku.com/heroku-linux-x64.tar.gz
unzip heroku-linux-x64.tar.gz
tar xf heroku-linux-x64.tar
sudo yum -y install glibc.i686
/home/ec2-user/heroku/bin/heroku --version
Once Heroku is installed, we need to set up a project for where we

will deploy projects. We can use the CLI to create a new Heroku
project by running the following commands:
/home/ec2-user/heroku/bin/heroku login
/home/ec2-user/heroku/bin/heroku create
This will create a unique app name, such as obscure-coast-69593. It’s

good to test the setup locally before deploying to production. In
order to test the setup, you’ll need to install the django and django-
heroku packages. Heroku has some dependencies on Postgres, which
is why additional install and easy_install commands are included
when installing these libraries.
pip install --user django

sudo yum install gcc python-setuptools postgresql-devel
sudo easy_install psycopg2
pip install --user django-heroku
To get started with building a Heroku application, we’ll first down-

load the sample application from GitHub and then modify the
project to include our echo application.
sudo yum install git

git clone https://github.com/heroku/python-getting-started.git
cd python-getting-started
2.4 Deploying a Web Endpoint 51
Next, we’ll make our changes to the project. We copy our echo.py
file into the directory, add Flask to the list of dependencies in the
requirements.txt file, override the command to run in the Procfile,
and then call heroku local to test the configuration locally.
cp ../echo.py echo.py
echo 'flask' >> requirements.txt
echo "web: gunicorn echo:app" > Procfile
/home/ec2-user/heroku/bin/heroku local
You should see a result that looks like this:
/home/ec2-user/heroku/bin/heroku local
[OKAY] Loaded ENV .env File as KEY=VALUE Format
[INFO] Starting gunicorn 19.9.0
[INFO] Listening at: http://0.0.0.0:5000 (10485)
[INFO] Using worker: sync
[INFO] Booting worker with pid: 10488
As before, we can test the endpoint using a browser or a Python

call, as shown below. In the Heroku local test configuration, port
5000 is used by default.
result = requests.get("http://localhost:5000/",
params = { 'msg': 'Hello from Heroku Local'})
The final step is to deploy the service to production. The git com-
mands are used to push the results to Heroku, which automatically
releases a new version of the application. The last command tells
Heroku to scale up to a single worker, which is free.
git add echo.py

git commit .
git push heroku master
/home/ec2-user/heroku/bin/heroku ps:scale web=1
After these steps run, there should be a message that the applica-
tion has been deployed to Heroku. Now we can call the endpoint,
which has a proper URL, is secured, and can be used to publicly
share data science projects.
result = requests.get("https://obscure-coast-69593.herokuapp.com",
params = { 'msg': 'Hello from Heroku Prod' })
There’s many languages and tools supported by Heroku, and it’s

useful for hosting small-scale data science projects.
2.5 Interactive Web Services

While the standard deployment of a model as a web service is an
API that you can call programmatically, it’s often useful to expose
models as interactive web applications. For example, we might
want to build an application where there is a UI for specifying
different inputs to a model, and the UI reacts to changes made by
the user. While Flask can be used to build web pages that react
to user input, there are libraries built on top of Flask that provide
higher-level abstractions for building web applications with the
Python language.
2.5.1 Dash
Dash is a Python library written by the Plotly team than enables
building interactive web applications with Python. You specify an
application layout and a set of callbacks that respond to user input.
If you’ve used Shiny in the past, Dash shares many similarities,
but is built on Python rather than R. With Dash, you can create
simple applications as we’ll show here, or complex dashboards that
interact with machine learning models.
We’ll create a simple Dash application that provides a UI for in-
teracting with a model. The application layout will contain three
2.5 Interactive Web Services 53
text boxes, where two of these are for user inputs and the third one
shows the output of the model. We’ll create a file called dash_app.py
and start by specifying the libraries to import.
import dash
import dash_html_components as html
import dash_core_components as dcc
from dash.dependencies import Input, Output
import pandas as pd
Next, we’ll define the layout of our application. We create a Dash

object and then set the layout field to include a title and three
text boxes with labels. We’ll include only 2 of the 10 games from
the games data set, to keep the sample short. The last step in
the script launches the web service and enables connections from
remote machines.
app = dash.Dash(__name__)
app.layout = html.Div(children=[
html.H1(children='Model UI'),
html.P([
html.Label('Game 1 '),
dcc.Input(value='1', type='text', id='g1'),
]),
html.Div([
html.Label('Game 2 '),
dcc.Input(value='0', type='text', id='g2'),
]),
html.P([
html.Label('Prediction '),
dcc.Input(value='0', type='text', id='pred')
]),
])
FIGURE 2.2: The initial Dash application.
if __name__ == '__main__':
app.run_server(host='0.0.0.0')
Before writing the callbacks, we can test out the layout of the
application by running python3 dash_app.py, which will run on port
8050 by default. You can browse to your public IP on port 8050
to see the resulting application. The initial application layout is
shown in Figure 2.2. Before any callbacks are added, the result of
the Prediction text box will always be 0.
The next step is to add a callback to the application so that the
Prediction text box is updated whenever the user changes one of
the Game 1 or Game 2 values. To perform this task, we define
a callback shown in the snippet below. The callback is defined
after the application layout, but before the run_server command.
We also load the logistic regression model for the games data set
using MLflow. The callback uses an annotation to define the inputs
to the function, the output, and any additional state that needs
to be provided. The way that the annotation is defined here, the
function will be called whenever the value of Game 1 or Game 2 is
modified by the user, and the value returned by this function will
be set as the value of the Prediction text box.
2.6 Interactive Web Services 55
model = mlflow.sklearn.load_model(model_path)
@app.callback(
Output(component_id='pred', component_property='value'),
[Input(component_id='g1', component_property='value'),
Input(component_id='g2', component_property='value')]
)
def update_prediction(game1, game2):
new_row = { "G1": float(game1),

"G2": float(game2),
"G3": 0, "G4": 0,
"G5": 0, "G6": 0,
"G7": 0, "G8": 0,
"G9": 0, "G10":0 }
return str(model.predict_proba(new_x)[0][1])
The function takes the two values provided by the user, and creates
a Pandas dataframe. As before, we transpose the dataframe to
provide a single row that we’ll pass as input to the loaded model.
The value predicted by the model is then returned and set as the
value of the Prediction text box.
The updated application with the callback function included
is shown in Figure 2.3. The prediction value now dynamically
changes in response to changes in the other text fields, and pro-
vides a way of introspecting the model.
Dash is great for building web applications, because it eliminates
the need to write JavaScript code. It’s also possible to stylize Dash
application using CSS to add some polish to your tools.
FIGURE 2.3: The resulting model prediction.
2.6 Conclusion
The Python ecosystem has a great suite of tools for building web
applications. Using only Python, you can write scalable APIs de-
ployed to the open web or custom UI applications that interact
with backend Python code. This chapter focused on Flask, which
can be extended with other libraries and hosted in a wide range
of environments. One of the important concepts we touched on
in this chapter is model persistence, which will be useful in other
contexts when building scalable model pipelines. We also deployed
a simple application to Heroku, which is a separate cloud platform
from AWS and GCP.
This chapter is only an introduction to the many different web
tools within the Python ecosystem, and the topic of scaling these
types of tools is outside the scope of this book. Instead, we’ll focus
on managed solutions for models on the web, which significantly
reduces the DevOps overhead of deploying models as web services.
The next chapter will cover two systems for serverless functions in
managed environments.
3
Models as Serverless Functions
Serverless technologies enable developers to write and deploy code

without needing to worry about provisioning and maintaining
servers. One of the most common uses of this technology is server-
less functions, which makes it much easier to author code that can
scale to match variable workloads. With serverless function envi-
ronments, you write a function that the runtime supports, specify
a list of dependencies, and then deploy the function to production.
The cloud platform is responsible for provisioning servers, scaling
up more machines to match demand, managing load balancers, and
handling versioning. Since we’ve already explored hosting models
as web endpoints, serverless functions are an excellent tool to uti-
lize when you want to rapidly move from prototype to production
for your predictive models.
Serverless functions were first introduced on AWS in 2015 and
GCP in 2016. Both of these systems provide a variety of triggers
that can invoke functions, and a number of outputs that the func-
tions can trigger in response. While it’s possible to use serverless
functions to avoid writing complex code for glueing different com-
ponents together in a cloud platform, we’ll explore a much nar-
rower use case in this chapter. We’ll write serverless functions that
are triggered by an HTTP request, calculate a propensity score for
the passed in feature vector, and return the prediction as JSON.
For this specific use case, GCP’s Cloud Functions are much eas-
ier to get up and running, but we’ll explore both AWS and GCP
solutions.
In this chapter, we’ll introduce the concept of managed services,
where the cloud platform is responsible for provisioning servers.
Next, we’ll cover hosting sklearn and Keras models with Cloud
57
58 3 Models as Serverless Functions
Functions. To conclude, we’ll show how to achieve the same re-

sult for sklearn models with Lambda functions in AWS. We’ll also
touch on model updates and access control.
3.1 Managed Services

Since 2015, there’s been a movement in cloud computing to transi-
tion developers away from manually provisioning servers to using
managed services that abstract away the concept of servers. The
main benefit of this new paradigm is that developers can write code
in a staging environment and then push code to production with
minimal concerns about operational overhead, and the infrastruc-
ture required to match the required workload can be automatically
scaled as needed. This enables both engineers and data scientists
to be more active in DevOps, because much of the operational
concerns of the infrastructure are managed by the cloud provider.
Manually provisioning servers, where you ssh into the machines to
set up libraries and code, is often referred to as hosted deployments,
versus managed solutions where the cloud platform is responsible
for abstracting away this concern from the user. In this book, we’ll
cover examples in both of these categories. Here are some of the
different use cases we’ll cover:
• Web Endpoints: Single EC2 instance (hosted) vs AWS
Lambda (managed).
• Docker: Single EC2 instance (hosted) vs ECS (managed).
• Messaging: Kafka (hosted) vs PubSub (managed).
This chapter will walk through the first use case, migrating web
endpoints from a single machine to an elastic environment. We’ll
also work through examples that thread this distinction, such as
deploying Spark environments with specific machine configurations
and manual cluster management.
Serverless technologies and managed services are a powerful tool
for data scientists, because they enable a single developer to build
3.2 Cloud Functions (GCP) 59
data pipelines that can scale to massive workloads. It’s a powerful

tool for data scientists to wield, but there are a few trade-offs to
consider when using managed services. Here are some of the main
issues to consider when deciding between hosted and managed
solutions:
• Iteration: Are you rapidly prototyping on a product or iterating
on a system in production?
• Latency: Is a multi-second latency acceptable for your SLAs?
• Scale: Can your system scale to match peak workload demands?
• Cost: Are you willing to pay more for serverless cloud costs?
At a startup, serverless technologies are great because you have
low-volume traffic and have the ability to quickly iterate and try
out new architectures. At a certain scale, the dynamics change
and the cost of using serverless technologies may be less appealing
when you already have in-house expertise for provisioning cloud
services. In my past projects, the top issue that was a concern was
latency, because it can impact customer experiences. In chapter 8,
we’ll touch on this topic, because managed solutions often do not
scale well to large streaming workloads.
Even if your organization does not use managed services in daily
operations, it’s a useful skill set to get hands on with as a data
scientist, because it means that you can separate model training
from model deployment issues. One of the themes in this book is
that models do not need to be complex, but it can be complex
to deploy models. Serverless functions are a great approach for
demonstrating the ability to serve models at scale, and we’ll walk
through two cloud platforms that provide this capability.
3.2 Cloud Functions (GCP)

Google Cloud Platform provides an environment for serverless
functions called Cloud Functions. The general concept with this
tool is that you can write code targeted for Flask, but leverage the
managed services in GCP to provide elastic computing for your
Python code. GCP is a great environment to get started with

serverless functions, because it closely matches standard Python
development ecosystems, where you specify a requirements file and
application code.
We’ll build scalable endpoints that serve both sklearn and Keras
models with Cloud Functions. There are a few issues to be aware
of when writing functions in this environment:
• Storage: Cloud Functions run in a read-only environment, but
you can write to the /tmp directory.
• Tabs: Spaces versus tabs can cause issues in Cloud Functions,
and if you are working in the web editor versus familiar tools like
Sublime Text, these can be difficult to spot.
• sklearn: When using a requirements file, it’s important to differ-
entiate between sklearn and scikit-learn based on your imports.
We’ll use sklearn in this chapter.
Cloud platforms are always changing, so the specific steps outlined
in this chapter may change based on the evolution of these plat-
forms, but the general approach for deploying functions should
apply throughout these updates. As always, the approach I advo-
cate for is starting with a simple example, and then scaling to
more complex solutions as needed. In this section, we’ll first build
an echo service and then explore sklearn and Keras models.
3.2.1 Echo Service

GCP provides a web interface for authoring Cloud Functions. This
UI provides options for setting up the triggers for a function, spec-
ifying the requirements file for a Python function, and authoring
the implementation of the Flask function that serves the request.
To start, we’ll set up a simple echo service that reads in a param-
eter from an HTTP request and returns the passed in parameter
as the result.
In GCP, you can directly set up a Cloud Function as an HTTP
endpoint without needing to configure additional triggers. To get
started with setting up an echo service, perform the following ac-
tions in the GCP console:
FIGURE 3.1: Creating a Cloud Function.
1. Search for “Cloud Function”

2. Click on “Create Function”
3. Select “HTTP” as the trigger
4. Select “Allow unauthenticated invocations”
5. Select “Inline Editor” for source code
6. Select Python 3.7 as the runtime
An example of this process is shown in Figure 3.1. After perform-

ing these steps, the UI will provide tabs for the main.py and re-
quirements.txt files. The requirements file is where we will specify
libraries, such as flask >= 1.1.1, and the main file is where we’ll
implement our function behavior.
We’ll start by creating a simple echo service that parses out the msg
parameter from the passed in request and returns this parameter
as a JSON response. In order to use the jsonify function we need
to include the flask library in the requirements file. The require-

ments.txt file and main.py files for the simple echo service are shown
in the snippet below. The echo function here is similar to the echo
service we coded in Section 2.1.1, the main distinction here is that
we are no longer using annotations to specify the endpoints and
allowed methods. Instead, these settings are now being specified
using the Cloud Functions UI.
# requirements.txt
flask
#main.py
def echo(request):
from flask import jsonify

params = request.get_json()
if "msg" in params:
data["response"] = str(params['msg'])
return jsonify(data)
We can deploy the function to production by performing the fol-

lowing steps:
1. Update “Function to execute” to “echo”

2. Click “Create” to deploy
Once the function has been deployed, you can click on the “Testing”
tab to check if the deployment of the function worked as intended.
You can specify a JSON object to pass to the function, and invoke
the function by clicking “Test the function”, as shown in Figure 3.2.
The result of running this test case is the JSON object returned
in the Output dialog, which shows that invoking the echo function
worked correctly.
FIGURE 3.2: Testing a Cloud Function.
Now that the function is deployed and we enabled unauthenticated

access to the function, we can call the function over the web using
Python. To get the URL of the function, click on the “trigger”
tab. We can use the requests library to pass a JSON object to the
serverless function, as shown in the snippet below.
import requests
result = requests.post(
"https://us-central1-gameanalytics.cloudfunctions.net/echo"
,json = { 'msg': 'Hello from Cloud Function' })
The result of running this script is that a JSON payload is returned

from the serverless function. The output from the call is the JSON
shown below.
{
'response': 'Hello from Cloud Function',
'success': True
}
We now have a serverless function that provides an echo service. In

order to serve a model using Cloud Functions, we’ll need to persist
the model specification somewhere that the serverless function can
access. To accomplish this, we’ll use Cloud Storage to store the
model in a distributed storage layer.
3.2.2 Cloud Storage (GCS)

GCP provides an elastic storage layer called Google Cloud Storage
(GCS) that can be used for distributed file storage and can also
scale to other uses such as data lakes. In this section, we’ll explore
the first use case of utilizing this service to store and retrieve files
for use in a serverless function. GCS is similar to AWS’s offering
called S3, which is leveraged extensively in the gaming industry to
build data platforms.
While GCP does provide a UI for interacting with GCS, we’ll
explore the command line interface in this section, since this ap-
proach is useful for building automated workflows. GCP requires
authentication for interacting with this service, please revisit sec-
tion 1.5.1 if you have not yet set up a JSON credentials file. In
order to interact with Cloud Storage using Python, we’ll also need
to install the GCS library, using the command shown below:
pip install --user google-cloud-storage

export GOOGLE_APPLICATION_CREDENTIALS=/home/ec2-user/dsdemo.json
Now that we have the prerequisite libraries installed and creden-

tials set up, we can interact with GCS programmatically using
Python. Before we can store a file, we need to set up a bucket on
GCS. A bucket is a prefix assigned to all files stored on GCS, and
each bucket name must be globally unique. We’ll create a bucket
name called dsp_model_store where we’ll store model objects. The

script below shows how to create a new bucket using the cre-
ate_bucket function and then iterate through all of the available
buckets using the list_buckets function. You’ll need to change the
bucket_name variable to something unique before running this script.
from google.cloud import storage

bucket_name = "dsp_model_store"
storage_client = storage.Client()
storage_client.create_bucket(bucket_name)
for bucket in storage_client.list_buckets():

print(bucket.name)
After running this code, the output of the script should be a sin-
gle bucket, with the name assigned to the bucket_name variable.
We now have a path on GCS that we can use for saving files:
gs://dsp_model_storage.
We’ll reuse the model we trained in Section 2.2.1 to deploy a lo-

gistic regression model with Cloud Functions. To save the file to
GCS, we need to assign a path to the destination, shown by the
bucket.blob command below and select a local file to upload, which
is passed to the upload function.
bucket = storage_client.get_bucket(bucket_name)
blob = bucket.blob("serverless/logit/v1")
blob.upload_from_filename("logit.pkl")
After running this script, the local file logit.pkl will now be avail-
able on GCS at the following location:
gs://dsp_model_storage/serverless/logit/v1/logit.pkl
While it’s possible to use URIs such as this directly to access files,
as we’ll explore with Spark in Chapter 6, in this section we’ll re-
trieve the file using the bucket name and blob path. The code
snippet below shows how to download the model file from GCS to
local storage. We download the model file to the local path of lo-
cal_logit.pkl and then load the model by calling pickle.load with
this path.
import pickle
blob.download_to_filename("local_logit.pkl")
model = pickle.load(open("local_logit.pkl", 'rb'))
model
We can now programmatically store model files to GCS using

Python and also retrieve them, enabling us to load model files
in Cloud Functions. We’ll combine this with the Flask examples
from the previous chapter to serve sklearn and Keras models as
Cloud Functions.
3.2.3 Model Function

We can now set up a Cloud Function that serves logistic regression
model predictions over the web. We’ll build on the Flask example
that we explored in Section 2.3.1 and make a few modifications
for the service to run on GCP. The first step is to specify the
required Python libraries that we’ll need to serve requests in the
requirements.txt file, as shown below. We’ll also need Pandas to
set up a dataframe for making the prediction, sklearn for applying
the model, and cloud storage for retrieving the model object from
GCS.
google-cloud-storage
sklearn
pandas
flask
The next step is to implement our model function in the main.py file.
A small change from before is that the params object is now fetched
using request.get_json() rather than flask.request.args. The main
change is that we are now downloading the model file from GCS
rather than retrieving the file directly from local storage, because
local files are not available when writing Cloud Functions with the
UI tool. An additional change from the prior function is that we
are now reloading the model for every request, rather than loading
the model file once at startup. In a later code snippet, we’ll show
how to use global objects to cache the loaded model.
def pred(request):
import pickle as pk
import sklearn
import pandas as pd

if "G1" in params:

# set up access to the GCS bucket

# download and load the model

blob.download_to_filename("/tmp/local_logit.pkl")
model = pk.load(open("/tmp/local_logit.pkl", 'rb'))
data["response"] = str(model.predict_proba(new_x)[0][1])
One note in the code snippet above is that the /tmp directory is
used to store the downloaded model file. In Cloud Functions, you
are unable to write to the local disk, with the exception of this
directory. Generally it’s best to read objects directly into memory
rather than pulling objects to local storage, but the Python library
for reading objects from GCS currently requires this approach.
For this function, we created a new Cloud Function named pred,
set the function to execute to pred, and deployed the function to
production. We can now call the function from Python, using the
same approach from 2.3.1 with a URL that now points to the
Cloud Function, as shown below:
import requests
result = requests.post(
"https://us-central1-gameanalytics.cloudfunctions.net/pred"
,json = { 'G1':'1', 'G2':'0', 'G3':'0', 'G4':'0', 'G5':'0'

,'G6':'0', 'G7':'0', 'G8':'0', 'G9':'0', 'G10':'0'})
The result of the Python web request to the function is a JSON

response with a response value and model prediction, shown below:
{
'response': '0.06745113592634559',
'success': True
}
In order to improve the performance of the function, so that it

takes milliseconds to respond rather than seconds, we’ll need to
cache the model object between runs. It’s best to avoid defining
variables outside of the scope of the function, because the server
hosting the function may be terminated due to inactivity. Global
variables are an execution to this rule, when used for caching ob-
jects between function invocations. This code snippet below shows
how a global model object can be defined within the scope of the
pred function to provide a persistent object across calls. During
the first function invocation, the model file will be retrieved from
GCS and loaded via pickle. During following runs, the model ob-
ject will already be loaded into memory, providing a much faster
response time.
model = None
def pred(request):
global model
if not model:
# download model from GCS

model = pk.load(open("/tmp/local_logit.pkl", 'rb'))
# apply model
Caching objects is important for authoring responsive models that

lazily load objects as needed. It’s also useful for more complex mod-
els, such as Keras which requires persisting a TensorFlow graph
between invocations.
3.2.4 Keras Model

Since Cloud Functions provide a requirements file that can be used
to add additional dependencies to a function, it’s also possible to
serve Keras models with this approach. We’ll be able to reuse most
of the code from the past section, and we’ll also use the Keras
and Flask approach introduced in Section 2.3.2. Given the size of
the Keras libraries and dependencies, we’ll need to upgrade the
memory available for the Function from 256 MB to 1GB. We also
need to update the requirements file to include Keras:
google-cloud-storage
tensorflow
keras
pandas
flask
The full implementation for the Keras model as a Cloud Function

is shown in the code snippet below. In order to make sure that the
TensorFlow graph used to load the model is available for future
invocations of the model, we use global variables to cache both
the model and graph objects. To load the Keras model, we need
to redefine the auc function that was used during model training,
which we include within the scope of the predict function. We
reuse the same approach from the prior section to download the
model file from GCS, but now use load_model from Keras to read
the model file into memory from the temporary disk location. The
result is a Keras predictive model that lazily fetches the model file
and can scale to meet variable workloads as a serverless function.
model = None
graph = None
def predict(request):
global model
global graph

import pandas as pd
import flask
import keras as k
from keras.models import load_model

k.backend.get_session().run(
return auc

# download model if not cached

if not model:
graph = tf.get_default_graph()
bucket_name = "dsp_model_store_1"
blob = bucket.blob("serverless/keras/v1")
blob.download_to_filename("/tmp/games.h5")
model = load_model('/tmp/games.h5',
custom_objects={'auc':auc})
# apply the model

if "G1" in params:
with graph.as_default():
data["response"]= str(model.predict_proba(new_x)[0][0])
To test the deployed model, we can reuse the Python web request
script from the prior section and replace pred with predict in the
request URL. We have now deployed a deep learning model to
production.
3.2.5 Access Control

The Cloud Functions we introduced in this chapter are open to
the web, which means that anyone can access them and poten-
tially abuse the endpoints. In general, it’s best not to enable unau-
thenticated access and instead lock down the function so that only
authenticated users and services can access them. This recommen-
dation also applies to the Flask apps that we deployed in the last
chapter, where it’s a best practice to restrict access to services that
can reach the endpoint using AWS private IPs.
There are a few different approaches for locking down Cloud Func-
tions to ensure that only authenticated users have access to the
functions. The easiest approach is to disable “Allow unauthenti-
cated invocations” in the function setup to prevent hosting the
function on the open web. To use the function, you’ll need to set
up IAM roles and credentials for the function. This process involves
a number of steps and may change over time as GCP evolves. In-
stead of walking through this process, it’s best to refer to the GCP
documentation1 .
Another approach for setting up functions that enforce authenti-
cation is by using other services within GCP. We’ll explore this
approach in Chapter 8, which introduces GCP’s PubSub system
for producing and consuming messages within GCP’s ecosystem.
3.2.6 Model Refreshes

We’ve deployed sklearn and Keras models to production using
Cloud Functions, but the current implementations of these func-
tions use static model files that will not change over time. It’s
usually necessary to make changes to models over time to ensure
that the accuracy of the models do not drift too far from expected
performance. There’s a few different approaches that we can take
to update the model specification that a Cloud Function is using:
1. Redeploy: Overwriting the model file on GCS and rede-

ploying the function will result in the function loading the
updated file.
2. Timeout: We can add a timeout to the function, where
the model is re-downloaded after a certain threshold of
time passes, such as 30 minutes.
3. New Function: We can deploy a new function, such as
pred_v2 and update the URL used by systems calling the
service, or use a load balancer to automate this process.
4. Model Trigger: We can add additional triggers to the
function to force the function to manually reload the
model.
1
https://cloud.google.com/functions/docs/securing/authenticating
While the first approach is the easiest to implement and can work
well for small-scale deployments, the third approach, where a load
balancer is used to direct calls to the newest function available
is probably the most robust approach for production systems. A
best practice is to add logging to your function, in order to track
predictions over time so that you can log the performance of the
model and identify potential drift.
3.3 Lambda Functions (AWS)

AWS also provides an ecosystem for serverless functions called
Lambda. AWS Lambda is useful for glueing different components
within an AWS deployment together, since it supports a rich set
of triggers for function inputs and outputs. While Lambda does
provide a powerful tool for building data pipelines, the current
Python development environment is a bit clunkier than GCP.
In this section we’ll walk through setting up an echo service and
an sklearn model endpoint with Lambda. We won’t cover Keras,
because the size of the library causes problems when deploying a
function with AWS. Unlike the past section where we used a UI
to define functions, we’ll use command line tools for providing our
function definition to Lambda.
3.3.1 Echo Function

For a simple function, you can use the inline code editor that
Lambda provides for authoring functions. You can create a new
function by performing the following steps in the AWS console:
1. Under “Find Services”, select “Lambda”

2. Select “Create Function”
3. Use “Author from scratch”
4. Assign a name (e.g. echo)
5. Select a Python runtime
6. Click “Create Function”
3.3 Lambda Functions (AWS) 75
After running these steps, Lambda will generate a file called

lambda_function.py. The file defines a function called lambda_handler
which we’ll use to implement the echo service. We’ll make a small
modification to the file, as shown below, which echoes the msg pa-
rameter as the body of the response object.
def lambda_handler(event, context):
return {
'statusCode': 200,
'body': event['msg']
}
Click “Save” to deploy the function and then “Test” to test the
file. If you use the default test parameters, then an error will be
returned when running the function, because no msg key is available
in the event object. Click on “Configure test event”, and define use
the following configuration:
{
"msg": "Hello from Lambda!"
}
After clicking on “Test”, you should see the execution results. The
response should be the echoed message with a status code of 200
returned. There’s also details about how long the function took to
execute (25.8ms), the billing duration (100ms), and the maximum
memory used (56 MB).
We have now a simple function running on AWS Lambda. For this
function to be exposed to external systems, we’ll need to set up an
API Gateway, which is covered in Section 3.3.3. This function will
scale up to meet demand if needed, and requires no server mon-
itoring once deployed. To setup a function that deploys a model,
we’ll need to use a different workflow for authoring and publishing
the function, because AWS Lambda does not currently support a
requirements.txt file for defining dependencies when writing func-
tions with the inline code editor. To store the model file that we
want to serve with a Lambda function, we’ll use S3 as a storage
layer for model artifacts.
3.3.2 Simple Storage Service (S3)

AWS provides a highly-performant storage layer called S3, which
can be used to host individual files for web sites, store large files
for data processing, and even host thousands or millions of files
for building data lakes. For now, our use case will be storing an
individual zip file, which we’ll use to deploy new Lambda functions.
However, there are many broader use cases and many companies
use S3 as their initial endpoint for data ingestion in data platforms.
In order to use S3 to store our function to deploy, we’ll need to
set up a new S3 bucket, define a policy for accessing the bucket,
and configure credentials for setting up command line access to S3.
Buckets on S3 are analogous to GCS buckets in GCP.
To set up a bucket, browse to the AWS console and select “S3” un-
der find services. Next, select “Create Bucket” to set up a location
for storing files on S3. Create a unique name for the S3 bucket, as
shown in Figure 3.3, and click “Next” and then “Create Bucket”
to finalize setting up the bucket.
We now have a location to store objects on S3, but we still need
to set up a user before we can use the command line tools to
write and read from the bucket. Browse to the AWS console and
select “IAM” under “Find Services”. Next, click “Users” and then
“Add user” to set up a new user. Create a user name, and select
“Programmatic access” as shown in Figure 3.4.
The next step is to provide the user with full access to S3. Use
the attach existing policies option and search for S3 policies in
order to find and select the AmazonS3FullAccess policy, as shown in
Figure 3.5. Click “Next” to continue the process until a new user
is defined. At the end of this process, a set of credentials will be
displayed, including an access key ID and secret access key. Store
these values in a safe location.
FIGURE 3.3: Creating an S3 bucket on AWS.
FIGURE 3.4: Setting up a user with S3 access.
The last step needed for setting up command line access to S3

is running the aws configure command from your EC2 instance.
You’ll be asked to provide the access and secret keys from the user
we just set up. In order to test that the credentials are properly
configured, you can run the following commands:
aws configure
aws s3 ls
FIGURE 3.5: Selecting a policy for full S3 access.
The results should include the name of the S3 bucket we set up

at the beginning of this section. Now that we have an S3 bucket
set up with command line access, we can begin writing Lambda
functions that use additional libraries such as Pandas and sklearn.
3.3.3 Model Function

In order to author a Lambda function that uses libraries outside of
the base Python distribution, you’ll need to set up a local environ-
ment that defines the function and includes all of the dependencies.
Once your function is defined, you can upload the function by cre-
ating a zip file of the local environment, uploading the resulting file
to S3, and configuring a Lambda function from the file uploaded
to S3.
The first step in this process is to create a directory with all of the
dependencies installed locally. While it’s possible to perform this
process on a local machine, I used an EC2 instance to provide a
clean Python environment. The next step is to install the libraries
needed for the function, which are Pandas and sklearn. These li-
braries are already installed on the EC2 instance, but need to be
reinstalled in the current directory in order to be included in the
zip file that we’ll upload to S3. To accomplish this, we can append
-t . to the end of the pip command in order to install the libraries
into the current directory. The last steps to run on the command
line are copying our logistic regression model into the current di-
rectory, and creating a new file that will implement the Lambda
function.
mkdir lambda
cd lambda
pip install pandas -t .
pip install sklearn -t .
cp ../logit.pkl logit.pkl
vi logit.py
The full source code for the Lambda function that serves our lo-
gistic regression model is shown in the code snippet below. The
structure of the file should look familiar, we first globally define a
model object and then implement a function that services model
requests. This function first parses the response to extract the in-
puts to the model, and then calls predict_proba on the resulting
dataframe to get a model prediction. The result is then returned
as a dictionary object containing a body key. It’s important to de-
fine the function response within the body key, otherwise Lambda
will throw an exception when invoking the function over the web.
from sklearn.externals import joblib

import pandas as pd
import json
model = joblib.load('logit.pkl')
def lambda_handler(event, context):
# read in the request body as the event dict

if "body" in event:
event = event["body"]
if event is not None:

event = json.loads(event)
else:
event = {}
if "G1" in event:
new_row = { "G1": event["G1"],"G2": event["G2"],
"G3": event["G3"],"G4": event["G4"],
"G5": event["G5"],"G6": event["G6"],
"G7": event["G7"],"G8": event["G8"],
"G9": event["G9"],"G10":event["G10"]}
prediction = str(model.predict_proba(new_x)[0][1])
return { "body": "Prediction " + prediction }
return { "body": "No parameters" }
Unlike Cloud Functions, Lambda functions authored in Python

are not built on top of the Flask library. Instead of requiring a
single parameter (request), a Lambda function requires event and
context objects to be passed in as function parameters. The event
includes the parameters of the request, and the context provides in-
formation about the execution environment of the function. When
testing a Lambda function using the “Test” functionality in the
Lambda console, the test configuration is passed directly to the
function as a dictionary in the event object. However, when the
function is called from the web, the event object is a dictionary
that describes the web request, and the request parameters are
stored in the body key in this dict. The first step in the Lambda
function above checks if the function is being called directly from
the console, or via the web. If the function is being called from
the web, then the function overrides the event dictionary with the
content in the body of the request.
One of the main differences from this approach with the GCP
Cloud Function is that we did not need to explicitly define global
variables that are lazily defined. With Lambda functions, you can
define variables outside the scope of the function that are persisted
before the function is invoked. It’s important to load model objects
outside of the model service function, because reloading the model
each time a request is made can become expensive when handling
large workloads.
To deploy the model, we need to create a zip file of the current
directory, and upload the file to a location on S3. The snippet
below shows how to perform these steps and then confirm that
the upload succeeded using the s3 ls command. You’ll need to
modify the paths to use the S3 bucket name that you defined in
the previous section.
zip -r logitFunction.zip .
aws s3 cp logitFunction.zip s3://dsp-ch3-logit/logitFunction.zip
aws s3 ls s3://dsp-ch3-logit/
Once your function is uploaded as a zip file to S3, you can return
to the AWS console and set up a new Lambda function. Select
“Author from scratch” as before, and under “Code entry type” se-
lect the option to upload from S3, specifying the location from the
cp command above. You’ll also need to define the Handler, which is
a combination of the Python file name and the Lambda function
name. An example configuration for the logit function is shown in
Figure 3.6.
Make sure to select the Python runtime as the same version of
Python that was used to run the pip commands on the EC2 in-
stance. Once the function is deployed by pressing “Save”, we can
test the function using the following definition for the test event.
{
"G1": "1", "G2": "1", "G3": "1",
"G4": "1", "G5": "1",
FIGURE 3.6: Defining the logit function on AWS Lambda.
"G6": "1", "G7": "1", "G8": "1",

"G9": "1", "G10": "1"
}
Since the model is loaded when the function is deployed, the re-
sponse time for testing the function should be relatively fast. An
example output of testing the function is shown in Figure 3.7. The
output of the function is a dictionary that includes a body key and
the output of the model as the value. The function took 110 ms
to execute and was billed for a duration of 200 ms.
So far, we’ve invoked the function only using the built-in test func-
tionality of Lambda. In order to host the function so that other
services can interact with the function, we’ll need to define an API
Gateway. Under the “Designer” tab, click “Add Trigger” and se-
lect “API Gateway”. Next, select “Create a new API” and choose
“Open” as the security setting. After setting up the trigger, an
API Gateway should be visible in the Designer layout, as shown
in Figure 3.8.
Before calling the function from Python code, we can use the API
Gateway testing functionality to make sure that the function is
set up properly. One of the challenges I ran into when testing this
FIGURE 3.7: Testing the logit function on AWS Lambda.
FIGURE 3.8: Setting up an API Gateway for the function.

FIGURE 3.9: Testing post commands on the Lambda function.
Lambda function was that the structure of the request varies when
the function is invoked from the web versus the console. This is
why the function first checks if the event object is a web request
or dictionary with parameters. When you use the API Gateway to
test the function, the resulting call will emulate calling the function
as a web request. An example test of the logit function is shown
in Figure 3.9.
Now that the gateway is set up, we can call the function from a
remote host using Python. The code snippet below shows how to
use a POST command to call the function and display the result.
Since the function returns a string for the response, we use the
text attribute rather than the json function to display the result.
import requests
result = requests.post("https://3z5btf0ucb.execute-api.us-east-1.
amazonaws.com/default/logit",
json = { 'G1':'1', 'G2':'0', 'G3':'0', 'G4':'0', 'G5':'0',
'G6':'0', 'G7':'0', 'G8':'0', 'G9':'0', 'G10':'0' })
print(result.text)
3.4 Conclusion 85
We now have a predictive model deployed to AWS Lambda that

will autoscale as necessary to match workloads, and which requires
minimal overhead to maintain.
Similar to Cloud Functions, there are a few different approaches
that can be used to update the deployed models. However, for
the approach we used in this section, updating the model requires
updating the model file in the development environment, rebuild-
ing the zip file and uploading it to S3, and then deploying a new
version of the model. This is a manual process and if you expect
frequent model updates, then it’s better to rewrite the function so
that it fetches the model definition from S3 directly rather than
expecting the file to already be available in the local context. The
most scalable approach is setting up additional triggers for the
function, to notify the function that it’s time to load a new model.
3.4 Conclusion
Serverless functions are a type of managed service that enable
developers to deploy production-scale systems without needing to
worry about infrastructure. To provide this abstraction, different
cloud platforms do place constraints on how functions must be
implemented, but the trade-off is generally worth the improvement
in DevOps that these tools enable. While serverless technologies
like Cloud Functions and Lambda can be operationally expensive,
they provide flexibility that can offset these costs.
In this chapter, we implemented echo services and sklearn model
endpoints using both GCP’s Cloud Functions and AWS’s Lambda
offerings. With AWS, we created a local Python environment with
all dependencies and then uploaded the resulting files to S3 to
deploy functions, while in GCP we authored functions directly
using the online code editor. The best system to use will likely
depend on which cloud provider your organization is already using,
but when prototyping new systems, it’s useful to have hands on
experience using more than one serverless function ecosystem.
4
Containers for Reproducible Models
When deploying data science models, it’s important to be able to

reproduce the same environment used for training and the environ-
ment used for serving. In Chapter 2, we used the same machine for
both environments, and in Chapter 3 we used a requirements.txt
file to ensure that the serverless ecosystem used for model serving
matched our development environment. Container systems such as
Docker provide a tool for building reproducible environments, and
they are much lighter weight than alternative approaches such as
virtual machines.
The idea of a container is that it is an isolated environment in
which you can set up the dependencies that you need in order to
perform a task. The task can be performing ETL work, serving
ML models, standing up APIs, or hosting interactive web appli-
cations. The goal of a container framework is to provide isolation
between instances with a lightweight footprint. With a container
framework, you specify the dependencies that your code needs, and
let the framework handle the legwork of managing different execu-
tion environments. Docker is the de facto standard for containers,
and there is substantial tooling built this platform.
Elastic container environments, such as Elastic Container Service
(ECS) provide similar functionality to serverless functions, where
you want to abstract away the notion of servers from hosting data
science models. The key differentiation is that serverless ecosys-
tems are restricted to specific runtimes, often have memory limi-
tations that make it challenging to use deep learning frameworks,
and are cloud specific. With ECS, you are responsible for setting
up the types of instances used to serve models, you can use what-
ever languages needed to serve the model, and you can take up
87
88 4 Containers for Reproducible Models
as much memory as needed. ECS still has the problem of being a

proprietary AWS tool, but newer options such as EKS build upon
Kubernetes which is open source and portable.
Here are some of the data science use cases I’ve seen for containers:
• Reproducible Analyses: Containers provide a great way of
packaging up analyses, so that other team members can rerun
your work months or years later.
• Web Applications: In Chapter 2 we built an interactive web
application with Dash. Containers provide a great way of ab-
stracting away hosting concerns for deploying the app.
• Model Deployments: If you want to expose your model as an
endpoint, containers provide a great way of separating the model
application code from model serving infrastructure.
The focus of this chapter will be the last use case. We’ll take our
web endpoint from Chapter 2 and wrap the application in a Docker
container. We’ll start by running the container locally on an EC2
instance, and then explore using ECS to create a scalable, load-
balanced, and fault-tolerant deployment of our model. We’ll then
show how to achieve a similar result on GCP using Kubernetes.
Now that we are exploring scalable compute environments, it’s
important to keep an eye on cloud costs when using ECS and GKE.
For AWS, it’s useful to keep an eye on how many EC2 instances are
provisioned, and on GCP the billing tool provides good tracking
of costs. The section on orchestration is specific to AWS and uses
an approach that is not portable to different cloud environments.
Feel free to skip directly to the section on Kubernetes if AWS is
not a suitable environment for your model deployments.
4.1 Docker
Docker, and other platform-as-a-service tools, provide a virtual-
ization concept called containers. Containers run on top of a host
operating system, but provide a standardized environment for code
4.1 Docker 89
running within the container. One of the key goals of this virtual-
ization approach is that you can write code for a target environ-
ment, and any system running Docker can run your container.
Containers are a lightweight alternative to virtual machines, which
provide similar functionality. The key difference is that containers
are much faster to spin up, while providing the same level of iso-
lation as virtual machines. Another benefit is that containers can
re-use layers from other containers, making it much faster to build
and share containers. Containers are a great solution to use when
you need to run conflicting versions of Python runtimes or libraries,
on a single machine.
With docker, you author a file called a Dockerfile that is used to
define the dependencies for a container. The result of building the
Dockerfile is a Docker Image, which packages all of the runtimes,
libraries, and code needed to run an app. A Docker Container
is an instantiated image that is running an application. One of
the useful features in Docker is that new images can build off
of existing images. For our model deployment, we’ll extend the
ubuntu:latest image.
This section will show how to set up Docker on an EC2 instance,

author a Dockerfile for building an image of the echo service from
Chapter 2, build an image using Docker, and run a container. To
install Docker on an EC2 instance, you can use the amazon-linux-
extras tool to simplify the process. The commands below will in-
stall Docker, start the service on the EC2 instance, and list the
running Containers, which will return an empty list.
sudo yum install -y python3-pip python3 python3-setuptools

sudo yum update -y
sudo amazon-linux-extras install docker
sudo service docker start
sudo docker ps
The application we’ll deploy is the echo service from Chapter 2.

This service is a Flask application that parses the msg attribute
from a GET or POST and returns a JSON payload echoing the

provided message. The only difference from the prior application
is that the Flask app now runs on port 80, shown by the last line
in the echo.py snippet below.
# load Flask
import flask
# define a predict function as an endpoint

@app.route("/predict", methods=["GET","POST"])
def predict():
# get the request parameters

params = flask.request.json
if (params == None):
# if parameters are found, echo the msg parameter

if (params != None):
data["response"] = params.get("msg")
# return a response in json format

# start the flask app, allow remote connections

app.run(host='0.0.0.0', port = 80)
Now that we have Docker installed and an application that we want

to containerize, we need to write a Dockerfile that describes how
to build an image. A Dockerfile for performing this task is shown
in the snippet below. The first step is to use the FROM command
to identify a base image to use. The ubuntu image provides a linux
environment that supports the apt-get command. The MAINTAINER
command adds to the metadata information associated with the
4.1 Docker 91
image, adding the name of the image maintainer. Next, the RUN
command is used to install Python, set up a symbolic link, and
install Flask. For containers with many Python libraries, it’s also
possible to use a requirements.txt file. The Copy command inserts
our script into the image and places the file in the root directory.
The final command specifies the arguments to run to execute the
application.
FROM ubuntu:latest
MAINTAINER Ben Weber
RUN apt-get update \

&& apt-get install -y python3-pip python3-dev \
&& cd /usr/local/bin \
&& ln -s /usr/bin/python3 python \
&& pip3 install flask
COPY echo.py echo.py
ENTRYPOINT ["python3","echo.py"]
After writing a Dockerfile, you can use the build command that
docker provides to create an image. The first command shown
in the snippet below shows how to build an image, tagged as
echo_service, using the file ./Dockerfile. The second command
shows the list of Docker images available on the instance. The
output will show both the ubuntu image we used as the base for
our image, and our newly created image.
sudo docker image build -t "echo_service" .

sudo docker images
To run an image as a container, we can use the run command

shown in the snippet below. The -d flag specifies that the container
should run as a daemon process, which will continue to run even
when shutting down the terminal. The -p flag is used to map a
port on the host machine to a port that the container uses for
communication. Without this setting, our container is unable to

receive external connections. The ps command shows the list of
running containers, which should now include the echo service.
sudo docker run -d -p 80:80 echo_service

sudo docker ps
To test the container, we can use the same process as before where
we use the external IP of the EC2 instance in a web browser and
pass a msg parameter to the /predict endpoint. Since we set up
a port mapping from the host port of 80 to the container port
80, we can directly invoke the container over the open web. An
example invocation and result from the echo service container is
shown below.
http://34.237.242.46/predict?msg=Hi_from_docker
{"response":"Hi_from_docker","success":true}
We have now walked through the process of building a Docker

image and running the image as a container on an EC2 instance.
While this approach does provide a solution for isolating differ-
ent services on a machine, it does not provide scaling and fault-
tolerance, which are typically requirements for a production-grade
model deployment.
4.2 Orchestration
Container orchestration systems are responsible for managing the
life cycles of containers in a cluster. They provide services including
provisioning, scaling, failover, load balancing, and service discovery
between containers. AWS has multiple orchestration solutions, but
the general trend has been moving towards Kubernetes for this
functionality, which is an open-source platform originally designed
by Google.
4.2 Orchestration 93
One of the main reasons for using container orchestration as a data

scientist is to be able to deploy models as containers, where you
can scale up infrastructure to match demand, have a fault-tolerant
system that can recover from errors, and have a static URL for
your service managed by a load balancer. It’s a bit of work to
get one of these solutions up and running, but the end result is
a robust model deployment. Serverless functions provide a similar
capability, but using orchestration solutions means that you can
use whatever programming language and runtime is necessary to
serve your models. It provides flexibility at the cost of operational
overhead, but as these tools evolve the overhead should be reduced.
The best solution to use for orchestration depends on your use case
and if you have constraints on the cloud platform that you can use.
If your goal is to use Kubernetes, then GCP provides a fully man-
aged solution out of the box, which we’ll cover in the next section.
If you are restricted to AWS, then the two options are Elastic
Container Service (ECS) and Elastic Kubernetes Service (EKS).
If you’re just getting started with Kubernetes, then AWS provides
less tools for getting containers up and running through a web con-
sole. Within ECS there are options for manually specifying EC2
instance types to run, or using the new Fargate mode to abstract
away managing EC2 instances. My general recommendation is to
learn Kubernetes if you want to get started with orchestration. But
if you need to run a containerized model at scale in AWS, then
ECS is currently the path of least friction.
In this section, we’ll walk through deploying our model image to
the AWS Container Registry, show how to set up a task in an ECS
cluster to run the image as a container, and then set up a load-
balanced service the provides managed provisioning for the echo
service.
4.2.1 AWS Container Registry (ECR)

In order to use your Docker image in an orchestration system, you
need to push your image to a Docker registry that works with
the platform. For ECS, the AWS implementation of this services
is called AWS Elastic Container Registry (ECR). ECR is a man-
FIGURE 4.1: A model repository on ECR.
aged Docker registry that you can use to store and manage images
within the AWS ecosystem. It works well with both ECS and EKS.
The goal of this subsection is to walk through the process of getting
a Docker image from an EC2 instance to ECR. We’ll cover the
following steps:
1. Setting up an ECR repository

2. Creating an IAM role for ECR
3. Using docker login
4. Tagging an image
5. Pushing an image
The first step is to create a repository for the image that we want
to store on ECR. A registry can have multiple repositories, and
each repository can have multiple tagged images. To set up a new
repository, perform the following steps from the AWS console:
1. Search for and select “ECR”

2. On the left panel, click on “Repositories”
3. Select “Create Repository”
4. Assign a repository name, e.g. “models”
5. Click on “Create Repository”
After completing these steps, you should have a new repository for
saving images on ECR, as shown in Figure 4.1. The repository will
initially be empty until we push a container.
Since our goal is to push a container from an EC2 instance to
ECR, we’ll need to set up permissions for pushing to the registry
from the command line. To set up these permissions, we can add

additional policies to the s3_lambda user that we first created in
Chapter 3. Perform the following steps from the AWS console:
1. Search for and select “IAM”

2. Click on Users
3. Select the “s3_lambda” user
4. Click “Add permissions”
5. Choose “AmazonEC2ContainerRegistryFullAccess”
6. Select “Attach existing policies directly”
7. Click on “Add permissions”
We now have permissions to write to ECR from the user account

we previously set up. However, before we can run the docker login
command for pushing images to ECR, we need to set up temporary
credentials for accessing ECR. We can create a token by running
the following command on the EC2 instance:
aws ecr get-login --region us-east-1 --no-include-email
The output of running this task is a command to run from the

command line that will enable temporary access for storing images
to ECR. For this to work on the EC2 instance, you’ll need to
prepend sudo to the output from the following step and then run
the generated command, as shown below.
sudo docker login -u AWS -p eyJwYXlsb2FkIjoiVy9vYWVrYnY0YlVqTFp...
If the authentication is successful, the output from this command

should be Login Succeeded. Now that we have successfully used the
docker login command, we can now push to the ECR repository.
Before we can push our local image, we need to tag the image with
account specific parameters. The snippet below shows how to tag
the echo service with an AWS account ID and region, with a tag
of echo. For these parameters, region is a string such as us-east-1,
and the account ID is the number specified under Account ID in the
“My Account” tab from the AWS console.
FIGURE 4.2: The echo image in the ECR repository.
sudo docker tag echo_service

[acount_id].dkr.ecr.[region].amazonaws.com/models:echo
After tagging your image, it’s good to check that the outcome
matches the expected behavior. To check the tags of your images,
run sudo docker images from the command line. An example output
is shown below, with my account ID and region omitted.
REPOSITORY TAG IMAGE ID CREATED SIZE

amazonaws.com/models echo 3380f2a8805b 3 hours ago 473MB
echo_service latest 3380f2a8805b 3 hours ago 473MB
ubuntu latest cf0f3ca922e0 8 days ago 64.2MB
The final step is to push the tagged image to the ECR repository.
We can accomplish this by running the command shown below:
sudo docker push [acount_id].dkr.[region].amazonaws.com/models:echo
After running this command, the echo service should now be avail-
able in the model repository on ECR. To check if the process suc-
ceeded, return the the AWS console and click on “Images” for the
model repository. The repo should now should an image with the
tag models:echo, as shown in Figure 4.2.
The outcome of this process is that we now have a Docker image
pushed to ECR that can be leveraged by an orchestration system.
While we did walk through a number of AWS specific steps, the

process of using Docker login applies to other cloud platforms.
4.2.2 AWS Container Service (ECS)

AWS provides an elastic container service called ECS that is a good
platform for getting started with container management. While it
is a proprietary technology, it does introduce many concepts that
are useful to consider in an orchestrated cloud deployment. Here
are some concepts exposed through an ECS deployment:
• Cluster: A cluster defines the environment used to provision
and host containers.
• Task: A task is an instantiation of a container than performs a
specific workload.
• Service: A service manages a task and provisions new machines
based on demand.
Now that we have an image hosted on ECR, the next step is to
use ECS to provide a scalable deployment of this image. In this
section, we’ll walk through the following steps:
1. Setting up a cluster
2. Setting up a task
3. Running a task
4. Running a service
At the end of this section, we’ll have a service that manages a task
running the echo service, but we’ll connect directly to the IP of
the provisioned EC2 instance. In the next section, we’ll set up a
load balancer to provide a static URL for accessing the service.
The first step in using ECS is to set up a service. There is a newer
feature in ECS called Fargate that abstracts away the notion of
EC2 instances when running a cluster, but this mode does not cur-
rently support the networking modes that we need for connecting
directly to the container. To set up an ECS cluster, perform the
following steps from the AWS console:
FIGURE 4.3: An empty ECS cluster.
1. Search for and select “ECS”

2. Click on “Clusters” on the left
3. Select “Create Cluster”
4. Use the “EC2 Linux + Networking” option
5. Assign a name, “dsp”
6. Select an instance type, m3.medium
7. For “VPC”, select an existing VPC
8. For “IAM Role”, use “aws-elasticbeanstalk-ec2-role”
9. Click “Create” to start the cluster
We now have an empty cluster with no tasks or services defined,

as shown in Figure 4.3. While the cluster has not yet spun up any
EC2 instances, there is an hourly billing amount associated with
running the cluster, and any EC2 instances that do spin up will
not be eligible for the free-tier option.
The next step is to define a task, which specifies the image to
use and a number of different settings. A task can be executed di-
rectly, or managed via a service. Tasks are independent of services
and a single task can be used across multiple services if necessary.
To set up a task, we’ll first specify the execution environment by

performing the following steps:
1. From the ECS console, click “Task Definitions”

2. Select “Create a new Task Definition”
3. Select EC2 launch type
4. Assign a name, “echo_task”
5. For “Task Role”, select “ecsTaskExecutionRole”
6. For “Network Mode”, use “Bridge”
7. Select “ecsTaskExecutionRole” for Task execution role
8. Set “Task Memory” and “Task CPU” to 1024
We then need to specify details about the Docker image that we

want to host when running the task:
1. Click “Add container”

2. Assign a name “echo_service”
3. For “Image”, use the URI shown in Figure 4.2
4. Add a port mapping of host:80, container:80
5. Click “Add” to finalize container setup
6. Click “Create” to define the task
We now have a task setup in ECS that we can use to host our
image as a container in the cloud. It is a good practice to test our
your tasks in ECS before defining a service to manage the task.
We can test out the task by performing the following steps:
1. From the ECS console, click on your cluster

2. Click on the “Tasks” tab
3. Select “Run new Task”
4. Use EC2 as the “Launch type”
5. For “Task Definition”, use “echo_task:1”
6. Click “Run Task”
The result of setting up and running the echo service as a task

is shown in Figure 4.4. The ECS cluster is now running a single
task, which is hosting our echo service. To provision this task, the
ECS cluster will spin up a new EC2 instance. If you browse to
FIGURE 4.4: The echo task running on the ECS cluster.
FIGURE 4.5: Network bindings for the echo service container.
the EC2 console in AWS, you’ll see that a new EC2 instance has
been provisioned, and the name of the instance will be based on
the service, such as: ECS Instance - EC2ContainerService-echo.
Now that we have a container running in our ECS cluster, we can
query it over the open web. To find the URL of the service, click
on the running task and under containers, expand the echo service
details. The console will show an external IP address where the
container can be accessed, as shown in Figure 4.5. An example of
using the echo service is shown in the snippet below.
http://18.212.21.97/predict?msg=Hi_from_ECS
{"response":"Hi_from_ECS","success":true}
We now have a container running in the cloud, but it’s not scalable
and there is no fault tolerance. To enable these types of capabilities
we need to define a service in our ECS cluster than manages the
FIGURE 4.6: The ECS service running a single echo task.
lifecycle of a task. To set up a service for the task, perform the

following steps:
1. From the ECS console, click on your cluster

2. Select the “Services” tab
3. Click “Create”
4. Use EC2 launch type
5. Select “echo_task” as the task
6. Assign a “Service name”, “echo_ecs”
7. Use “None” for load balancer type
8. Click “Create Service”
This will start the service. The service will set the “Desired count”
value to 1, and it may take a few minutes for a new task to get
ramped up by the cluster. Once “Running count” is set to 1, you
can start using the service to host a model. An example of the
provisioned service is shown in Figure 4.6. To find the IP of the
container, click on the task within the service definition.
We now have a container that is managed by a service, but we’re
still accessing the container directly. In order to use ECS in a way
FIGURE 4.7: The Application Load Balancer configuration.
that scales, we need to set up a load balancer that provides a static

URL and routes requests to active tasks managed by the service.
4.2.3 Load Balancing

There’s a number of different load balancer options in AWS that
are useful for different deployments. This is also an area where the
options are rapidly changing. For our ECS cluster, we can use the
application load balancer to provide a static URL for accessing
the echo service. To set up a load balancer, perform the following
steps from the AWS console:
1. Search for and select “EC2”

2. Select “Load Balancer” on the left
3. Click “Create Load Balancer”
4. Choose “Application Load Balancer”
5. Assign a name, “model-service”
6. Use the default VPC
7. Create a new security group
8. Set source to “Anywhere”
9. Create a new target group, “model-group”
An example of a provisioned application load balancer is shown in

Figure 4.7. The last step needed to set up a load-balanced container
is configuring an ECS service that uses the load balancer. Repeat
the same steps shown in the prior section for setting up a service,
but instead of selecting “None” for the load balancer type, perform
the following actions:
1. Select “Application” for Load Balance type

2. Select the “model-service” balancer
3. Select the “echo_service” container
4. Click “Add to Load Balancer”
5. Select “model-group” as the target group
6. Click “Create Service”
It’s taken quite a few steps, but our echo service is now running in
a scalable environment, using a load balancer, and using a service
that will manage tasks to handle failures and provision new EC2
instances as necessary. This approach is quite a bit of configuration
versus Lambda for similar functionality, but this approach may be
preferred based on the type of workload that you need to handle.
There is cost involved with running an ECS cluster, even if you
are not actively servicing requests, so understanding your expected
workload is useful when modeling out the cost of different model
serving options on AWS.
http://model123.us-east-1.elb.amazonaws.com/predict?msg=Hi_from_ELB
{"response":"Hi_from_ELB","success":true}
AWS does provide an option for Kubernetes called EKS, but the
options available through the web console are currently limited for
managing Docker images. EKS can work with ECR as well, and
as the AWS platform evolves EKS will likely be the best option
for new deployments.
Make sure to terminate your cluster, load balancers, and EC2 in-
stances once you are done testing out your deployment to reduce
your cloud platform costs.
4.3 Kubernetes on GCP

Google Cloud Platform provides a service called Google Kuber-
netes Engine (GKE) for serving Docker containers. Kubernetes is
a container-orchestration system originally developed by Google
that is now open source. There are a wide range of use cases for
this platform, but we’ll focus on the specific task of hosting our
echo service using managed Kubernetes.
Using Kubernetes for hosting a Docker container is similar to ECS,
where the first step is to save your image to a Docker registry that
can interface with the orchestration system. The GCP version of
this registry service is called Container Registry. To get our im-
age from an EC2 instance on AWS to the GCP Container Registry,
we’ll again use the docker login command. For this process to work,
you’ll need the GCP credentials json file that we set up in Chap-
ter 1. The code snippet below shows how to pass the json file to
the docker login command, tag the image for uploading it to the
registry, and push the image to Container Registry.
cat dsdemo.json | sudo docker login -u _json_key

--password-stdin https://us.gcr.io
sudo docker tag echo_service us.gcr.io/[gcp_account]/echo_service
sudo docker push us.gcr.io/[gcp_account]/echo_service
You’ll need to replace the gcp_acount parameter in this script with

your full google account ID. After performing these steps, the echo
service image should be visible under the Registry view in the GCP
console, as shown in Figure 4.8. Typically, if you are using GCP
for serving models, it’s likely that you’ll be using Google Compute
instances rather than EC2, but it’s good to get practice interfacing
between components in different cloud platforms.
The process for hosting a container with GKE is streamlined com-
pared to all of the steps needed for using ECS. We’ll first use the
GCP console to set up a container on Kubernetes, and then expose
4.3 Kubernetes on GCP 105
FIGURE 4.8: The echo image on GCP Container Registry.
FIGURE 4.9: The echo image deployed via Kubernetes.
the service to the open web. To deploy the echo service container,
perform the following steps from the GCP console:
1. Search for and select “Kubernetes Engine”

2. Click “Deploy Container”
3. Select “Existing Container Image”
4. Choose “echo_service:latest”
5. Assign an application name “echo-kge”
6. Click “Deploy”
We now have a Kubernetes cluster deployed and ready to serve the

echo service. The deployment of a Kubernetes cluster on GKE can
take a few minutes to set up. Once the deployment is completed,
you should see the echo cluster under the list of clusters, as shown
in Figure 4.9.
FIGURE 4.10: The echo service deployed to the open web.
To use the service, we’ll need to expose the cluster to the open
web by performing the following steps from the GCP console:
1. From the GKE menu, select your cluster

2. Click on “Workloads”
3. Choose the “echo-gke” workload
4. Select the “Actions” tab and then “Expose”
5. For service type, select “load balancer”
After performing these steps, the cluster will configure an external

IP that can be used to invoke the service, as shown in Figure
4.10. GKE will automatically load balance and scale the service as
needed to match the workload.
http://35.238.43.63/predict?msg=Hi_from_GKE
{"response":"Hi_from_GKE","success":true}
An example of using the service is shown in the snippet above.

We were able to quickly take a Docker image and deploy it in a
Kubernetes ecosystem using GKE. It’s good to build experience
with Kubernetes for hosting Docker images, because it is a portable
4.4 Conclusion 107
solution that works across multiple cloud environments and it is

being adopted by many open-source projects.
4.4 Conclusion
Containers are great to use to make sure that your analyses and
models are reproducible across different environments. While con-
tainers are useful for keeping dependencies clean on a single ma-
chine, the main benefit is that they enable data scientists to write
model endpoints without worrying about how the container will
be hosted. This separation of concerns makes it easier to partner
with engineering teams to deploy models to production, or using
the approaches shown in this chapter, data and applied science
teams can also own the deployment of models to production.
The best approach to use for serving models depends on your
deployment environment and expected workload. Typically, you
are constrained to a specific cloud platform when working at a
company, because your model service may need to interface with
other components in the cloud, such as a database or cloud stor-
age. Within AWS, there are multiple options for hosting contain-
ers while GCP is aligned on GKE as a single solution. The main
question to ask is whether it is more cost effective to serve your
model using serverless function technologies or elastic container
technologies. The correct answer will depend on the volume of
traffic you need to handle, the amount of latency that is tolera-
ble for end users, and the complexity of models that you need to
host. Containerized solutions are great for serving complex models
and making sure that you can meet latency requirements, but may
require a bit more DevOps overhead versus serverless functions.
5
Workflow Tools for Model Pipelines
Model pipelines are usually part of a broader data platform that

provides data sources, such as lakes and warehouses, and data
stores, such as an application database. When building a pipeline,
it’s useful to be able to schedule a task to run, ensure that any de-
pendencies for the pipeline have already completed, and to backfill
historic data if needed. While it’s possible to perform these types
of tasks manually, there are a variety of tools that have been de-
veloped to improve the management of data science workflows.
In this chapter, we’ll explore a batch model pipeline that per-
forms a sequence of tasks in order to train and store results for
a propensity model. This is a different type of task than the de-
ployments we’ve explored so far, which have focused on serving
real-time model predictions as a web endpoint. In a batch process,
you perform a set of operations that store model results that are
later served by a different application. For example, a batch model
pipeline may predict which users in a game are likely to churn, and
a game server fetches predictions for each user that starts a session
and provides personalized offers.
When building batch model pipelines for production systems, it’s
important to make sure that issues with the pipeline are quickly
resolved. For example, if the model pipeline is unable to fetch the
most recent data for a set of users due to an upstream failure
with a database, it’s useful to have a system in place that can
send alerts to the team that owns the pipeline and that can rerun
portions of the model pipeline in order to resolve any issues with
the prerequisite data or model outputs.
Workflow tools provide a solution for managing these types of
problems in model pipelines. With a workflow tool, you specify
109
110 5 Workflow Tools for Model Pipelines
the operations that need to be completed, identify dependencies

between the operations, and then schedule the operations to be
performed by the tool. A workflow tool is responsible for running
tasks, provisioning resources, and monitoring the status of tasks.
There’s a number of open source tools for building workflows in-
cluding Airflow, Luigi, MLflow, and Pentaho Kettle. We’ll focus
on Airflow, because it is being widely adopted across companies
and cloud platforms and are also providing fully-managed versions
of Airflow.
In this chapter, we’ll build a batch model pipeline that runs as a
Docker container. Next, we’ll schedule the task to run on an EC2
instance using cron, and then explore a managed version of cron
using Kubernetes. In the third section, we’ll use Airflow to define a
graph of operations to perform in order to run our model pipeline,
and explore a cloud offering of Airflow.
5.1 Sklearn Workflow

A common workflow for batch model pipelines is to extract data
from a data lake or data warehouse, train a model on historic user
behavior, predict future user behavior for more recent data, and
then save the results to a data warehouse or application database.
In the gaming industry, this is a workflow I’ve seen used for build-
ing likelihood to purchase and likelihood to churn models, where
the game servers use these predictions to provide different treat-
ments to users based on the model predictions. Usually libraries
like sklearn are used to develop models, and languages such as
PySpark are used to scale up to the full player base.
It is typical for model pipelines to require other ETLs to run in a
data platform before the pipeline can run on the most recent data.
For example, there may be an upstream step in the data platform
that translates json strings into schematized events that are used
as input for a model. In this situation, it might be necessary to
rerun the pipeline on a day that issues occurred with the json
5.1 Sklearn Workflow 111
transformation process. For this section, we’ll avoid this complica-

tion by using a static input data source, but the tools that we’ll
explore provide the functionality needed to handle these issues.
There’s typically two types of batch model pipelines that can I’ve
seen deployed in the gaming industry:
• Persistent: A separate training workflow is used to train models
from the one used to build predictions. A model is persisted
between training runs and loaded in the serving workflow.
• Transient: The same workflow is used for training and serving
predictions, and instead of saving the model as a file, the model
is rebuilt for each run.
In this section we’ll build a transient batch pipeline, where a new
model is retrained with each run. This approach generally results
in more compute resources being used if the training process is
heavyweight, but it helps avoid issues with model drift, which is
useful to track. We’ll author a pipeline that performs the following
steps:
1. Fetches a data set from GitHub

2. Trains a logistic regression model
3. Applies the regression model
4. Saves the results to BigQuery
The pipeline will execute as a single Python script that performs all
of these steps. For situations where you want to use intermediate
outputs from steps across multiple tasks, it’s useful to decompose
the pipeline into multiple processes that are integrated through a
workflow tool such as Airflow.
We’ll build this workflow by first writing a Python script that runs
on an EC2 instance, and then Dockerize the script so that we can
use the container in workflows. To get started, we need to install
a library for writing a Pandas dataframe to BigQuery:
pip install --user pandas_gbq

Next, we’ll create a file called pipeline.py that performs the four
pipeline steps identified above. The script shown below performs
these steps by loading the necessary libraries, fetching the CSV
file from GitHub into a Pandas dataframe, splits the dataframe
into train and test groups to simulate historic and more recent
users, builds a logistic regression model using the training data
set, creates predictions for the test data set, and saves the resulting
dataframe to BigQuery.
import pandas as pd
import numpy as np
from google.oauth2 import service_account
from datetime import datetime
import pandas_gbq
# fetch the data set and add IDs

gamesDF['User_ID'] = gamesDF.index
gamesDF['New_User'] = np.floor(np.random.randint(0, 10,
gamesDF.shape[0])/9)
# train and test groups

train = gamesDF[gamesDF['New_User'] == 0]
x_train = train.iloc[:,0:10]
y_train = train['label']
test = gameDF[gamesDF['New_User'] == 1]
x_test = test.iloc[:,0:10]
# build a model
y_pred = model.predict_proba(x_test)[:, 1]
# build a predictions dataframe

resultDF = pd.DataFrame({'User_ID':test['User_ID'], 'Pred':y_pred})

resultDF['time'] = str(datetime. now())
# save predictions to BigQuery

table_id = "dsp_demo.user_scores"
project_id = "gameanalytics-123"
credentials = service_account.Credentials.
from_service_account_file('dsdemo.json')
pandas_gbq.to_gbq(resultDF, table_id, project_id=project_id,
if_exists = 'replace', credentials=credentials)
To simulate a real-world data set, the script assigns a User_ID at-

tribute to each record, which represents a unique ID to track differ-
ent users in a system. The script also splits users into historic and
recent groups by assigning a New_User attribute. After building pre-
dictions for each of the recent users, we create a results dataframe
with the user ID, the model prediction, and a timestamp. It’s use-
ful to apply timestamps to predictions in order to determine if the
pipeline has completed successfully. To test the model pipeline,
run the following statements on the command line:
export GOOGLE_APPLICATION_CREDENTIALS=
/home/ec2-user/dsdemo.json
python3 pipeline.py
If successfully, the script should create a new data set on BigQuery

called dsp_demo, create a new table called user_users, and fill the
table with user predictions. To test if data was actually populated
in BigQuery, run the following commands in Jupyter:

sql = "select * from dsp_demo.user_scores"

client.query(sql).to_dataframe().head()
FIGURE 5.1: Querying the uploaded predictions in BigQuery.
This script will set up a client for connecting to BigQuery and then
display the result set of the query submitted to BigQuery. You can
also browse to the BigQuery web UI to inspect the results of the
pipeline, as shown in Figure 5.1. We now have a script that can
fetch data, apply a machine learning model, and save the results
as a single process.
With many workflow tools, you can run Python code or bash
scripts directly, but it’s good to set up isolated environments for
executing scripts in order to avoid dependency conflicts for differ-
ent libraries and runtimes. Luckily, we explored a tool for this in
Chapter 4 and can use Docker with workflow tools. It’s useful to
wrap Python scripts in Docker for workflow tools, because you can
add libraries that may not be installed on the system responsible
for scheduling, you can avoid issues with Python version conflicts,
and containers are becoming a common way of defining tasks in
workflow tools.
To containerize our workflow, we need to define a Dockerfile, as
shown below. Since we are building out a new Python environ-
ment from scratch, we’ll need to install Pandas, sklearn, and the
BigQuery library. We also need to copy credentials from the EC2
instance into the container so that we can run the export com-
mand for authenticating with GCP. This works for short term
deployments, but for longer running containers it’s better to run
the export in the instantiated container rather than copying static
credentials into images. The Dockerfile lists out the Python li-
braries needed to run the script, copies in the local files needed for
execution, exports credentials, and specifies the script to run.
FROM ubuntu:latest
MAINTAINER Ben Weber
RUN apt-get update \

&& apt-get install -y python3-pip python3-dev \
&& cd /usr/local/bin \
&& ln -s /usr/bin/python3 python \
&& pip3 install pandas \
&& pip3 install sklearn \
&& pip3 install pandas_gbq
COPY pipeline.py pipeline.py

COPY /home/ec2-user/dsdemo.json dsdemo.json
RUN export GOOGLE_APPLICATION_CREDENTIALS=/dsdemo.json
ENTRYPOINT ["python3","pipeline.py"]
Before deploying this script to production, we need to build an

image from the script and test a sample run. The commands below
show how to build an image from the Dockerfile, list the Docker
images, and run an instance of the model pipeline image.
sudo docker image build -t "sklearn_pipeline" .

sudo docker images
sudo docker run sklearn_pipeline
After running the last command, the containerized pipeline should

update the model predictions in BigQuery. We now have a model
pipeline that we can run as a single bash command, which we
now need to schedule to run at a specific frequency. For testing

purposes, we’ll run the script every minute, but in practice models
are typically executed hourly, daily, or weekly.
5.2 Cron
A common requirement for model pipelines is running a task at
a regular frequency, such as every day or every hour. Cron is a
utility that provides scheduling functionality for machines running
the Linux operating system. You can Set up a scheduled task using
the crontab utility and assign a cron expression that defines how
frequently to run the command. Cron jobs run directly on the
machine where cron is utilized, and can make use of the runtimes
and libraries installed on the system.
There are a number of challenges with using cron in production-
grade systems, but it’s a great way to get started with scheduling
a small number of tasks and it’s good to learn the cron expression
syntax that is used in many scheduling systems. The main issue
with the cron utility is that it runs on a single machine, and does
not natively integrate with tools such as version control. If your
machine goes down, then you’ll need to recreate your environment
and update your cron table on a new machine.
A cron expression defines how frequently to run a command. It is
a sequence of 5 numbers that define when to execute for different
time granularities, and it can include wildcards to always run for
certain time periods. A few sample expressions are shown in the
snippet below:
# run every minute

* * * * *
# Run at 10am UTC everyday

0 10 * * *
5.2 Cron 117
# Run at 04:15 on Saturday

15 4 * * 6
When getting started with cron, it’s good to use tools1 to validate
your expressions. Cron expressions are used in Airflow and many
other scheduling systems.
We can use cron to schedule our model pipeline to run on a reg-
ular frequency. To schedule a command to run, run the following
command on the console:
crontab -e
This command will open up the cron table file for editing in vi.
To schedule the pipeline to run every minute, add the following
commands to the file and save.
# run every minute

* * * * * sudo docker run sklearn_pipeline
After exiting the editor, the cron table will be updated with the
new command to run. The second part of the cron statement is the
command to run. When defining the command to run, it’s useful
to include full file paths. With Docker, we just need to define the
image to run. To check that the script is actually executing, browse
to the BigQuery UI and check the time column on the user_scores
model output table.
We now have a utility for scheduling our model pipeline on a regu-
lar schedule. However, if the machine goes down then our pipeline
will fail to execute. To handle this situation, it’s good to explore
cloud offerings with cron scheduling capabilities.
1
https://crontab.guru/
5.2.1 Cloud Cron

Both AWS and GCP provide managed cron options. With AWS
it’s possible to schedule services to run on ECS using cron ex-
pressions, and the GCP Kubernetes Engine provides scheduling
support for containers. In this section, we’ll explore the GKE op-
tion because it is simpler to set up. The first step is to push our
model pipeline image to Container Registry using the following
commands:
cat dsdemo.json | sudo docker login -u _json_key

--password-stdin https://us.gcr.io
sudo docker tag sklearn_pipeline
us.gcr.io/[gcp_account]/sklearn_pipeline
sudo docker push us.gcr.io/[gcp_account]/sklearn_pipeline
Next, we’ll set up Kubernetes to schedule the image to run on

a managed cluster. Unlike Section 4.3 where we created a clus-
ter from an image directly, we’ll use Kubernetes control (kubectl)
commands to set up a scheduled container. Run the following com-
mands from the GCP console to create a GKE cluster:
1. Search for and select “Kubernetes Engine”

2. Click “Create Cluster”
3. Select “Your first cluster”
To schedule a task on this cluster, we’ll use a YAML file to con-

figure our task and then use a kubectl command to update the
cluster with the model pipeline. Perform the following steps from
the GKE UI to get access to the cluster and run a file editor:
1. Select the new cluster

2. Click on “Connect”
3. Select “Run in Cloud Shell”
4. Run the generated gcloud command
5. Run vi sklearn.yaml
5.2 Cron 119
This will provide terminal access to the console and allow us to

save a YAML file on the cluster. The YAML file below shows
how to define a task to run on a schedule with the specified cron
expression.
apiVersion: batch/v1beta1
kind: CronJob
metadata:
name: sklearn
spec:
schedule: "* * * * *"
jobTemplate:
spec:
template:
spec:
containers:
- name: sklearn
image: us.gcr.io/[gcp_account]/sklearn_pipeline
restartPolicy: OnFailure
After saving the file, we can use kubectl to update the cluster with
the YAML file. Run the command below to update the cluster
with the model pipeline task:
bgweber@cloudshell:~ (dsp)$ kubectl apply -f sklearn.yaml

cronjob.batch/sklearn created
To view the task, click on “workloads”. The events tab provides a

log of the runs of the task, as shown in Figure 5.2. Again, we can
validate that the pipeline is running successfully by browsing to
the BigQuery UI and checking the time attribute of predictions.
There’s a variety of scheduling options for cloud platforms, and we
explored just one option with GKE. GCP also provides a service
called Cloud Scheduler, but this system does not work directly
with Container Registry. Kubernetes is a good approach for han-
dling scheduling in a cloud deployment, because the system is also
FIGURE 5.2: Execution events for the pipeline task on GKE.
responsible for provisioning machines that the scheduled task will

use to execute.
5.3 Workflow Tools

Cron is useful for simple pipelines, but runs into challenges when
tasks have dependencies on other tasks which can fail. To help re-
solve this issue, where tasks have dependencies and only portions
of a pipeline need to be rerun, we can leverage workflow tools.
Apache Airflow is currently the most popular tool, but other open
source projects are available and provide similar functionality in-
cluding Luigi and MLflow.
There are a few situations where workflow tools provide benefits
over using cron directly:
• Dependencies: Workflow tools define graphs of operations,
which makes dependencies explicit.
• Backfills: It may be necessary to run an ETL on old data, for
a range of different dates.
• Versioning: Most workflow tools integrate with version control
systems to manage graphs.
• Alerting: These tools can send out emails or generate PageDuty
alerts when failures occur.
5.3 Workflow Tools 121
Workflow tools are particularly useful in environments where dif-

ferent teams are scheduling tasks. For example, many game com-
panies have data scientists that schedule model pipelines which
are dependent on ETLs scheduled by a seperate engineering team.
In this section, we’ll schedule our task to run an EC2 instance
using hosted Airflow, and then explore a fully-managed version of
Airflow on GCP.
5.3.1 Apache Airflow

Airflow is an open source workflow tool that was originally devel-
oped by Airbnb and publically released in 2015. It helps solve a
challenge that many companies face, which is scheduling tasks that
have many dependencies. One of the core concepts in this tool is
a graph that defines the tasks to perform and the relationships
between these tasks.
In Airflow, a graph is referred to as a DAG, which is an acronym
for directed acyclic graph. A DAG is a set of tasks to perform,
where each task has zero or more upstream dependencies. One of
the constraints is that cycles are not allowed, where two tasks have
upstream dependencies on each other.
DAGs are set up using Python code, which is one of the differences
from other workflow tools such as Pentaho Kettle which is GUI
focused. The Airflow approach is called “configuration as code”,
because a Python script defines the operations to perform within
a workflow graph. Using code instead of a GUI to configure work-
flows is useful because it makes it much easier to integrate with
version control tools such as GitHub.
To get started with Airflow, we need to install the library, initialize
the service, and run the scheduler. To perform these steps, run the
following commands on an EC2 instance or your local machine:
export AIRFLOW_HOME=~/airflow
pip install --user apache-airflow
FIGURE 5.3: The Airflow web app running on an EC2 instance.
airflow initdb
airflow scheduler
Airflow also provides a web frontend for managing DAGs that have
been scheduled. To start this service, run the following command
in a new terminal on the same machine.
airflow webserver -p 8080
This command tells Airflow to start the web service on port 8080.
You can open a web browser at this port on your machine to view
the web frontend for Airflow, as shown in Figure 5.3.
Airflow comes preloaded with a number of example DAGs. For our
model pipeline we’ll create a new DAG and then notify Airflow of
the update. We’ll create a file called sklearn.py with the following
DAG definition:
from airflow import DAG

from airflow.operators.bash_operator import BashOperator
from datetime import datetime, timedelta
default_args = {
'owner': 'Airflow',
'depends_on_past': False,
'email': '[email protected]',
'start_date': datetime(2019, 11, 1),
'email_on_failure': True,
}
dag = DAG('games', default_args=default_args,

schedule_interval="* * * * *")
t1 = BashOperator(
task_id='sklearn_pipeline',
bash_command='sudo docker run sklearn_pipeline',
dag=dag)
There’s a few steps in this Python script to call out. The script
uses a Bash operator to define the action to perform. The Bash
operator is defined as the last step in the script, which specifies
the operation to perform. The DAG is instantiated with a number
of input arguments that define the workflow settings, such as who
to email when the task fails. A cron expression is passed to the
DAG object to define the schedule for the task, and the DAG
object is passed to the Bash operator to associate the task with
this graph of operations.
Before adding the DAG to airflow, it’s useful to check for syntax
errors in your code. We can run the following command from the
terminal to check for issues with the DAG:
python3 sklearn.py
This command will not run the DAG, but will flag any syntax
errors present in the script. To update Airflow with the new DAG
file, run the following command:
airflow list_dags
-------------------------------------------------------------------
FIGURE 5.4: The sklearn DAG scheduled on Airflow.
DAGS
-------------------------------------------------------------------
games
This command will add the DAG to the list of workflows in Airflow.
To view the list of DAGs, navigate to the Airflow web server, as
shown in Figure 5.4. The web server will show the schedule of the
DAG, and provide a history of past runs of the workflow. To check
that the DAG is actually working, browse to the BigQuery UI and
check for fresh model outputs.
We now have an Airflow service up and running that we can use
to monitor the execution of our workflows. This setup enables us
to track the execution of workflows, backfill any gaps in data sets,
and enable alerting for critical workflows.
Airflow supports a variety of operations, and many companies au-
thor custom operators for internal usage. In our first DAG, we used
the Bash operator to define the task to execute, but other options
are available for running Docker images, including the Docker op-
erator. The code snippet below shows how to change our DAG to
use the Docker operator instead of the Bash operator.
from airflow.operators.docker_operator import DockerOperator
t1 = DockerOperator(
image='sklearn_pipeline',
dag=dag)
The DAG we defined does not have any dependencies, since the
container performs all of the steps in the model pipeline. If we
had a dependency, such as running a sklearn_etl container before
running the model pipeline, we can use the set_upstrean command
as shown below. This configuration sets up two tasks, where the
pipeline task will execute after the etl task completes.
t1 = BashOperator(
task_id='sklearn_etl',
bash_command='sudo docker run sklearn_etl',
dag=dag)
t2 = BashOperator(
bash_command='sudo docker run sklearn_pipeline',
dag=dag)
t2.set_upstream(t1)
Airflow provides a rich set of functionality and we’ve only touched

the surface of what the tool provides. While we were already able
to schedule the model pipeline with hosted and managed cloud
offerings, it’s useful to schedule the task through Airflow for im-
proved monitoring and versioning. The landscape of workflow tools
will change over time, but many of the concepts of Airflow will
translate to these new tools.
5.3.2 Managed Airflow

We now have a workflow tool set up for managing model workflows,
but the default configuration does not provide high-availability,
where new machines are provisioned when failures occur. While it’s
possible to set up a distributed version of Airflow using Celery to
set up different Scheduler and Worker nodes, one of the recent trends
is using Kubernetes to create more robust Airflow deployments.
It is possible to self-host Airflow on Kubernetes, but it can be
complex to set up. There are also fully-managed versions of Airflow
available for cloud platforms such as Cloud Composer on GCP.
With a managed version of Airflow, you define the DAGs to execute
and set the schedules, and the platform is responsible for providing
a high-availability deployment.
To run our DAG on Cloud Composer, we’ll need to update the
task to use a GKE Pod operator in place of a Docker operator, be-
cause Composer needs to be able to authenticate with Container
Registry. The updated DAG is shown in the snippet below.
from airflow.gcp.operators.kubernetes_engine import GKEPodOperator
t1 = GKEPodOperator(
project_id = '{your_project_id}',
cluster_name = ' us-central1-models-13d59d5b-gke',
name ='sklearn_pipeline',
namespace='default',
location='us-central1-c',
image='us.gcr.io/{your_project_id}/sklearn_pipeline',
dag=dag)
Cloud Composer is built on top of GKE. A beta version was re-

leased in 2018 and many features of the tool are still evolving.
To get started with Composer, perform the following steps in the
GCP Console:
1. Browse to Cloud Composer2

3. Assign a name, “models”
4. Select Python 3
2
https://console.cloud.google.com/composer/
5.4 Conclusion 127
FIGURE 5.5: Fully-managed Airflow on Cloud Composer.
5. Select The most recent image version

Like GKE, cluster setup takes quite a while to complete. Once

setup completes, you can access the Airflow web service by clicking
on “Airflow” in the list of clusters, as shown in Figure 5.5.
Before adding our DAG, we’ll need to update the cluster_name at-
tribute to point to the provisioned GKE cluster. To find the clus-
ter name, click on the Composer cluster and find the GKE cluster
attribute. To add the DAG, click on “DAGs” in the cluster list,
which will direct your browser to a google storage bucket. After
uploading the DAG to this bucket using the upload UI, the Air-
flow cluster will automatically detect the new DAG and add it
to the list of workflows managed by this cluster. We now have a
high-availability Airflow deployment running our model workflow.
5.4 Conclusion
In this chapter we explored a batch model pipeline for applying a
machine learning model to a set of users and storing the results to
BigQuery. To make the pipeline portable, so that we can execute

it in different environments, we created a Docker image to define
the required libraries and credentials for the pipeline. We then
ran the pipeline on an EC2 instance using batch commands, cron,
and Airflow. We also used GKE and Cloud Composer to run the
container via Kubernetes.
Workflow tools can be tedious to set up, especially when installing
a cluster deployment, but they provide a number of benefits over
manual approaches. One of the key benefits is the ability to han-
dle DAG configuration as code, which enables code reviews and
version control for workflows. It’s useful to get experience with
configuration as code, because more and more projects are using
this approach.
6
PySpark for Batch Pipelines
Spark is a general-purpose computing framework that can scale

to massive data volumes. It builds upon prior big data tools such
as Hadoop and MapReduce, while providing significant improve-
ments in the expressivity of the languages it supports. One of the
core components of Spark is resilient distributed datasets (RDD),
which enable clusters of machines to perform workloads in a coordi-
nated, and fault-tolerant manner. In more recent versions of Spark,
the Dataframe API provides an abstraction on top of RDDs that
resembles the same data structure in R and Pandas. PySpark is
the Python interface to Spark, and it provides an API for working
with large-scale datasets in a distributed computing environment.
PySpark is an extremely valuable tool for data scientists, because
it can streamline the process for translating prototype models into
production-grade model workflows. At Zynga, our data science
team owns a number of production-grade systems that provide
useful signals to our game and marketing teams. By using PyS-
park, we’ve been able to reduce the amount of support we need
from engineering teams to scale up models from concept to pro-
duction.
Up until now in this book, all of the models we’ve built and de-
ployed have been targeted at single machines. While we are able
to scale up model serving to multiple machines using Lambda,
ECS, and GKS, these containers worked in isolation and there was
no coordination among nodes in these environments. With PyS-
park, we can build model workflows that are designed to operate
in cluster environments for both model training and model serving.
The result is that data scientists can now tackle much larger-scale
problems than previously possible using prior Python tools. PyS-
129
130 6 PySpark for Batch Pipelines
park provides a nice tradeoff between an expressive programming

language and APIs to Spark versus more legacy options such as
MapReduce. A general trend is that the use of Hadoop is drop-
ping as more data science and engineering teams are switching to
Spark ecosystems. In Chapter 7 we’ll explore another distributed
computing ecosystem for data science called Cloud Dataflow, but
for now Spark is the open-source leader in this space. PySpark was
one of the main motivations for me to switch from R to Python
for data science workflows.
The goal of this chapter is to provide an introduction to PySpark
for Python programmers that shows how to build large-scale model
pipelines for batch scoring applications, where you may have bil-
lions of records and millions of users that need to be scored. While
production-grade systems will typically push results to application
databases, in this chapter we’ll focus on batch processes that pull
in data from a lake and push results back to the data lake for other
systems to use. We’ll explore pipelines that perform model appli-
cations for both AWS and GCP. While the data sets used in this
chapter rely on AWS and GCP for storage, the Spark environment
does not have to run on either of these platforms and instead can
run on Azure, other clouds, or on-pem Spark clusters.
We’ll cover a variety of different topics in this chapter to show
different use cases of PySpark for scalable model pipelines. After
showing how to make data available to Spark on S3, we’ll cover
some of the basics of PySpark focusing on dataframe operations.
Next, we’ll build out a predictive model pipeline that reads in
data from S3, performs batch model predictions, and then writes
the results to S3. We’ll follow this by showing off how a newer
feature called Pandas UDFs can be used with PySpark to perform
distributed deep learning and feature engineering. To conclude,
we’ll build another batch model pipeline now using GCP and then
discuss how to productize workflows in a Spark ecosystem.
6.1 Spark Environments 131
6.1 Spark Environments

There’s a variety of ways to both configure Spark clusters and
submit commands to a cluster for execution. When getting started
with PySpark as a data scientist, my recommendation is to use a
freely-available notebook environment for getting up and running
with Spark as quick as possible. While PySpark may not perform
quite as well as Java or Scala for large-scale workflows, the ease of
development in an interactive programming environment is worth
the trade-off.
Based on your organization, you may be starting from scratch for
Spark or using an existing solution. Here are the types of Spark
deployments I’ve seen in practice:
• Self Hosted: An engineering team manages a set of clusters and
provides console and notebook access.
• Cloud Solutions: AWS provides a managed Spark option called
EMR and GCP has Cloud Dataproc.
• Vendor Solutions: Databricks, Cloudera, and other vendors
provide fully-managed Spark environments.
There’s a number of different factors to consider when choosing a
Spark ecosystem, including cost, scalability, and feature sets. As
you scale the size of the team using Spark, additional consider-
ations are whether an ecosystem supports multi-tenancy, where
multiple jobs can run concurrently on the same cluster, and isola-
tion where one job failing should not kill other jobs. Self-hosted
solutions require significant engineering work to support these ad-
ditional considerations, so many organizations use cloud or vendor
solutions for Spark. In this book, we’ll use the Databricks Com-
munity Edition, which provides all of the baseline features needed
for learning Spark in a collaborative notebook environment.
Spark is a rapidly evolving ecosystem, and it’s difficult to author
books about this subject that do not quickly become out of date
as the platform evolves. Another issue is that many books target
Scala rather than Python for the majority of coding examples. My
advice for readers that want to dig deeper into the Spark ecosystem
is to explore books based on the broader Spark ecosystem, such
as (Karau et al., 2015). You’ll likely need to read through Scala or
Java code examples, but the majority of content covered will be
relevant to PySpark.
6.1.1 Spark Clusters

A Spark environment is a cluster of machines with a single driver
node and zero or more worker nodes. The driver machine is the
master node in the cluster and is responsible for coordinating the
workloads to perform. In general, workloads will be distributed
across the worker nodes when performing operations on Spark
dataframes. However, when working with Python objects, such
as lists or dictionaries, objects will be instantiated on the driver
node.
Ideally, you want all of your workloads to be operating on worker
nodes, so that the execution of the steps to perform is distributed
across the cluster, and not bottlenecked by the driver node. How-
ever, there are some types of operations in PySpark where the
driver has to perform all of the work. The most common situa-
tion where this happens is when using Pandas dataframes in your
workloads. When you use toPandas or other commands to convert a
data set to a Pandas object, all of the data is loaded into memory
on the driver node, which can crash the driver node when working
with large data sets.
In PySpark, the majority of commands are lazily executed, mean-
ing that an operation is not performed until an output is explic-
itly needed. For example, a join operation between two Spark
dataframes will not immediately cause the join operation to be per-
formed, which is how Pandas works. Instead, the join is performed
once an output is added to the chain of operations to perform,
such as displaying a sample of the resulting dataframe. One of
the key differences between Pandas operations, where operations
are eagerly performed and pulled into memory, is that PySpark
operations are lazily performed and not pulled into memory until
needed. One of the benefits of this approach is that the graph of
operations to perform can be optimized before being sent to the

cluster to execute.
In general, nodes in a Spark cluster should be considered
ephemeral, because a cluster can be resized during execution. Addi-
tionally, some vendors may spin up a new cluster when scheduling
a job to run. This means that common operations in Python, such
as saving files to disk, do not map directly to PySpark. Instead,
using a distributed computing environment means that you need
to use a persistent file store such as S3 when saving data. This
is important for logging, because a worker node may crash and
it may not be possible to ssh into the node for debugging. Most
Spark deployments have a logging system set up to help with this
issue, but it’s good practice to log workflow status to persistent
storage.
6.1.2 Databricks Community Edition

One of the quickest ways to get up and running with PySpark is
to use a hosted notebook environment. Databricks is the largest
Spark vendor and provides a free version for getting started called
Community Edition1 . We’ll use this environment to get started
with Spark and build AWS and GCP model pipelines.
The first step is to create a login on the Databricks website for the
community edition. Next, perform the following steps to spin up
a test cluster after logging in:
1. Click on “Clusters” on the left navigation bar

3. Assign a name, “DSP”
4. Select the most recent runtime (non-beta)
After a few minutes we’ll have a cluster set up that we can use
for submitting Spark commands. Before attaching a notebook to
the cluster, we’ll first set up the libraries that we’ll use throughout
1
https://community.cloud.databricks.com/
this chapter. Instead of using pip to install libraries, we’ll use the
Databricks UI, which makes sure that every node in the cluster
has the same set of libraries installed. We’ll use both Maven and
PyPI to install libraries on the cluster. To install the BigQuery
connector, perform the following steps:
1. Click on “Clusters” on the left navigation bar

2. Select the “DSP” cluster
3. Click on the “Libraries” tab
4. Select “Install New”
5. Click on the “Maven” tab.
6. Set coordinates to com.spotify:spark-bigquery_2.11:0.2.2
7. Click install
The UI will then show the status as resolving, and then installing,
and then installed. We also need to attach a few Python libraries
that are not pre-installed on a new Databricks cluster. Standard
libraries such as Pandas are installed, but you might need to up-
grade to a more recent version since the libraries pre-installed by
Databricks can lag significantly.
To install a Python library on Databricks, perform the same steps
as before up to step 5. Next, instead of selecting “Maven” choose
“PyPI”. Under Package, specify the package you want to install and
then click “Install”. To follow along with all of the sections in this
chapter, you’ll need to install the following Python packages:
• koalas - for dataframe conversion.
• featuretools - for feature generation.
• tensorflow - for a deep learning backend.
• keras - for a deep learning model.
You’ll now have a cluster set up capable of performing distributed
feature engineering and deep learning. We’ll start with basic Spark
commands, show off newer functionality such as the Koalas library,
and then dig into these more advanced topics. After this setup,
your cluster library setup should look like Figure 6.1. To ensure
that everything is set up successfully, restart the cluster and check
the status of the installed libraries.
FIGURE 6.1: Libraries attached to a Databricks cluster.
Now that we have provisioned a cluster and set up the required

libraries, we can create a notebook to start submitting commands
to the cluster. To create a new notebook, perform the following
steps:
1. Click on “Databricks” on the left navigation bar

2. Under “Common Tasks”, select “New Notebook”
3. Assign a name “CH6”
4. Select “Python” as the language
5. select “DSP” as the cluster
The result will be a notebook environment where you can start

running Python and PySpark commands, such as print("Hello
World!"). An example notebook running this command is shown
in Figure 6.2. We now have a PySpark environment up and run-
ning that we can use to build distributed model pipelines.
FIGURE 6.2: Running a Python command in Databricks.
6.2 Staging Data

Data is essential for PySpark workflows. Spark supports a variety
of methods for reading in data sets, including connecting to data
lakes and data warehouses, as well as loading sample data sets from
libraries, such as the Boston housing data set. Since the theme of
this book is building scalable pipelines, we’ll focus on using data
layers that work with distributed workflows. To get started with
PySpark, we’ll stage input data for a model pipeline on S3, and
then read in the data set as a Spark dataframe.
This section will show how to stage data to S3, set up credentials
for accessing the data from Spark, and fetching the data from S3
into a Spark dataframe. The first step is to set up a bucket on S3
for storing the data set we want to load. To perform this step, run
the following operations on the command line.
aws s3api create-bucket --bucket dsp-ch6 --region us-east-1

aws s3 ls
After running the command to create a new bucket, we use the ls

command to verify that the bucket has been successfully created.
Next, we’ll download the games data set to the EC2 instance and
then move the file to S3 using the cp command, as shown in the
snippet below.
6.2 Staging Data 137
wget https://github.com/bgweber/Twitch/raw/master/
Recommendations/games-expand.csv
aws s3 cp games-expand.csv s3://dsp-ch6/csv/games-expand.csv
In addition to staging the games data set to S3, we’ll also copy a
subset of the CSV files from the Kaggle NHL data set, which we
set up in Section 1.5.2. Run the following commands to stage the
plays and stats CSV files from the NHL data set to S3.
aws s3 cp game_plays.csv s3://dsp-ch6/csv/game_plays.csv

aws s3 cp game_skater_stats.csv
s3://dsp-ch6/csv/game_skater_stats.csv
aws s3 ls s3://dsp-ch6/csv/
We now have all of the data sets needed for the code examples
in this chapter. In order to read in these data sets from Spark,
we’ll need to set up S3 credentials for interacting with S3 from the
Spark cluster.
6.2.1 S3 Credentials
For production environments, it is better to use IAM roles to man-
age access instead of using access keys. However, the community
edition of Databricks constrains how much configuration is allowed,
so we’ll use access keys to get up and running with the examples
in this chapter. We already set up a user for accessing S3 from an
EC2 instance. To create a set of credentials for accessing S3 pro-
grammatically, perform the following steps from the AWS console:
1. Search for and select “IAM”

2. Click on “Users”
3. Select the user created in Section 3.3.2, “S3_Lambda”
4. Click “Security Credentials”
5. Click “Create Access Key”
The result will be an access key and a secret key enabling access
to S3. Save these values in a secure location, as we’ll use them in
the notebook to connect to the data sets on S3. Once you are done
with this chapter, it is recommended to revoke these credentials.
Now that we have credentials set up for access, we can return to
the Databricks notebook to read in the data set. To enable access
to S3, we need to set the access key and secret key in the Hadoop
configuration of the cluster. To set these keys, run the PySpark
commands shown in the snippet below. You’ll need to replace the
access and secret keys with the credentials we just created for the
S3_Lambda role.
AWS_ACCESS_KEY = "AK..."
AWS_SECRET_KEY = "dC..."
sc._jsc.hadoopConfiguration().set(
"fs.s3n.awsAccessKeyId", AWS_ACCESS_KEY)
"fs.s3n.awsSecretAccessKey", AWS_SECRET_KEY)
We can now read the data set into a Spark dataframe using the read
command, as shown below. This command uses the spark context
to issue a read command and reads in the data set using the CSV
input reader. We also specify that the CSV file includes a header
row and that we want Spark to infer the data types for the columns.
When reading in CSV files, Spark eagerly fetches the data set into
memory, which can cause issues for larger data sets. When working
with large CSV files, it’s a best practice to split up large data sets
into multiple files and then read in the files using a wildcard in
the input path. When using other file formats, such as Parquet or
Avro, Spark lazily fetches the data sets.
games_df = spark.read.csv("s3://dsp-ch6/csv/games-expand.csv",
header=True, inferSchema = True)
display(games_df)
The display command in the snippet above is a utility function pro-

vided by Databricks that samples the input dataframe and shows
6.3 A PySpark Primer 139
FIGURE 6.3: Displaying the dataframe in Databricks.
a table representation of the frame, as shown in Figure 6.3. It is

similar to the head function in Pandas, but provides additional
functionality such as transforming the sampled dataframe into a
plot. We’ll explore the plotting functionality in Section 6.3.3.
Now that we have data loaded into a Spark dataframe, we can be-
gin exploring the PySpark language, which enables data scientists
to build production-grade model pipelines.
6.3 A PySpark Primer

PySpark is a powerful language for both exploratory analysis and
building machine learning pipelines. The core data type in PyS-
park is the Spark dataframe, which is similar to Pandas dataframes,
but is designed to execute in a distributed environment. While the
Spark Dataframe API does provide a familiar interface for Python
programmers, there are significant differences in the way that com-
mands issued to these objects are executed. A key difference is that
Spark commands are lazily executed, which means that commands
such as iloc are not available on these objects. While working with
Spark dataframes can seem constraining, the benefit is that PyS-
park can scale to much larger data sets than Pandas.
This section will walk through common operations for Spark

dataframes, including persisting data, converting between different
dataframe types, transforming dataframes, and using user-defined
functions. We’ll use the NHL stats data set, which provides user-
level summaries of player performance for each game. To load this
data set as a Spark dataframe, run the commands in the snippet
below.
stats_df = spark.read.csv("s3://dsp-ch6/csv/game_skater_stats.csv",
display(stats_df)
6.3.1 Persisting Dataframes

A common operation in PySpark is saving a dataframe to persis-
tent storage, or reading in a data set from a storage layer. While
PySpark can work with databases such as Redshift, it performs
much better when using distributed file stores such as S3 or GCS.
In this chapter we’ll use these types of storage layers as the outputs
of model pipelines, but it’s also useful to stage data to S3 as inter-
mediate steps within a workflow. For example, in the AutoModel2
system at Zynga, we stage the output of the feature generation
step to S3 before using MLlib to train and apply machine learning
models.
The data storage layer to use will depend on your cloud platform.
For AWS, S3 works well with Spark for distributed data reads and
writes. When using S3 or other data lakes, Spark supports a variety
of different file formats for persisting data. Parquet is typically the
industry standard when working with Spark, but we’ll also explore
Avro and ORC in addition to CSV. Avro is a better format for
streaming data pipelines, and ORC is useful when working with
legacy data pipelines.
To show the range of data formats supported by Spark, we’ll take
the stats data set and write it to Avro, then Parquet, then ORC,
2
https://www.gamasutra.com/blogs/BenWeber/20190426/340293/
and finally CSV. After performing this round trip of data IO, we’ll
end up with our initial Spark dataframe. To start, we’ll save the
stats dataframe in Avro format, using the code snippet shown be-
low. This code writes the dataframe to S3 in Avro format using
the Databricks Avro writer, and then reads in the results using
the same library. The result of performing these steps is that we
now have a Spark dataframe pointing to the Avro files on S3. Since
PySpark lazily evaluates operations, the Avro files are not pulled
to the Spark cluster until an output needs to be created from this
data set.
# AVRO write
avro_path = "s3://dsp-ch6/avro/game_skater_stats/"
stats_df.write.mode('overwrite').format(
"com.databricks.spark.avro").save(avro_path)
# AVRO read
avro_df = sqlContext.read.format(
"com.databricks.spark.avro").load(avro_path)
Avro is a distributed file format that is record based, while the

Parquet and ORC formats are column based. It is useful for the
streaming workflows that we’ll explore in Chapter 8, because it
compresses records for distributed data processing. The output of
saving the stats dataframe in Avro format is shown in the snip-
pet below, which shows a subset of the status files and data files
generated when persisting a dataframe to S3 as Avro. Like most
scalable data formats, Avro will write records to several files, based
on partitions if specified, in order to enable efficient read and write
operations.
aws s3 ls s3://dsp-ch6/avro/game_skater_stats/
2019-11-27 23:02:43 1455 _committed_1588617578250853157
2019-11-27 22:36:31 1455 _committed_1600779730937880795
2019-11-27 23:02:40 0 _started_1588617578250853157
2019-11-27 23:31:42 0 _started_6942074136190838586
2019-11-27 23:31:47 1486327 part-00000-tid-6942074136190838586-

c6806d0e-9e3d-40fc-b212-61c3d45c1bc3-15-1-c000.avro
2019-11-27 23:31:43 44514 part-00007-tid-6942074136190838586-
c6806d0e-9e3d-40fc-b212-61c3d45c1bc3-22-1-c000.avro
Parquet on S3 is currently the standard approach for building

data lakes on AWS, and tools such as Delta Lake are leveraging
this format to provide highly-scalable data platforms. Parquet is
a columnar-oriented file format that is designed for efficient reads
when only a subset of columns are being accessed for an operation,
such as when using Spark SQL. Parquet is a native format for
Spark, which means that PySpark has built-in functions for both
reading and writing files in this format.
An example of writing the stats dataframe as Parquet files and
reading in the result as a new dataframe is shown in the snippet
below. In this example, we haven’t set a partition key, but as with
Avro, the dataframe will be split up into multiple files in order
to support highly-performant read and write operations. When
working with large-scale data sets, it’s useful to set partition keys
for the file export using the repartition function. After this section,
we’ll use Parquet as the primary file format when working with
Spark.
# parquet out
parquet_path = "s3a://dsp-ch6/games-parquet/"
avro_df.write.mode('overwrite').parquet(parquet_path)
# parquet in
parquet_df = sqlContext.read.parquet(parquet_path)
ORC is a another columnar format that works well with Spark.

The main benefit over Parquet is that it can support improved
compression, at the cost of additional compute cost. I’m including
it in this chapter, because some legacy systems still use this format.
An example of writing the stats dataframe to ORC and reading the
results back into a Spark dataframe is shown in the snippet below.
Like the Avro format, the ORC write command will distribute the
dataframe to multiple files based on the size.
# orc out
orc_path = "s3a://dsp-ch6/games-orc/"
parquet_df.write.mode('overwrite').orc(orc_path)
# orc in
orc_df = sqlContext.read.orc(orc_path)
To complete our round trip of file formats, we’ll write the results
back to S3 in the CSV format. To make sure that we write a single
file rather than a batch of files, we’ll use the coalesce command
to collect the data to a single node before exporting it.This is a
command that will fail with large data sets, and in general it’s
best to avoid using the CSV format when using Spark. However,
CSV files are still a common format for sharing data, so it’s useful
to understand how to export to this format.
# CSV out
csv_path = "s3a://dsp-ch6/games-csv-out/"
orc_df.coalesce(1).write.mode('overwrite').format(
"com.databricks.spark.csv").option("header","true").save(csv_path)
# and CSV read to finish the round trip

csv_df = spark.read.csv(csv_path, header=True, inferSchema = True)
The resulting dataframe is the same as the dataframe that we first

read in from S3, but if the data types are not trivial to infer, then
the CSV format can cause problems. When persisting data with
PySpark, it’s best to use file formats that describe the schema of
the data being persisted.
6.3.2 Converting Dataframes

While it’s best to work with Spark dataframes when authoring
PySpark workloads, it’s often necessary to translate between dif-
ferent formats based on your use case. For example, you might
need to perform a Pandas operation, such as selecting a specific
element from a dataframe. When this is required, you can use the
toPandas function to pull a Spark dataframe into memory on the
driver node. The PySpark snippet below shows how to perform this
task, display the results, and then convert the Pandas dataframe
back to a Spark dataframe. In general, it’s best to avoid Pandas
when authoring PySpark workflows, because it prevents distribu-
tion and scale, but it’s often the best way of expressing a command
to execute.
stats_pd = stats_df.toPandas()
stats_df = sqlContext.createDataFrame(stats_pd)
To bridge the gap between Pandas and Spark dataframes,

Databricks introduced a new library called Koalas that resembles
the Pandas API for Spark-backed dataframes. The result is that
you can author Python code that works with Pandas commands
that can scale to Spark-scale data sets. An example of converting
a Spark dataframe to Koalas and back to Spark is shown in the
following snippet. After converting the stats dataframe to Koalas,
the snippet shows how to calculate the average time on ice as well
as index into the Koalas frame. The intent of Koalas is to provide
a Pandas interface to Spark dataframes, and as the Koalas library
matures more Python modules may be able to take advantage of
Spark. The output from the snippet shows that the average time
on ice was 993 seconds per game for NHL players.
import databricks.koalas as ks
stats_ks = stats_df.to_koalas()
stats_df = stats_ks.to_spark()
print(stats_ks['timeOnIce'].mean())
print(stats_ks.iloc[:1, 1:2])
During the development of this book, Koalas is still preliminary

and only partially implemented, but it strives to provide a familiar
interface for Python coders. Both Pandas and Spark dataframes
can work with Koalas, and the snippet below shows how to go
from Spark to Koalas to Pandas to Spark, and Spark to Pandas
to Koalas to Spark.
# spark -> koalas -> pandas -> spark

df = sqlContext.createDataFrame(stats_df.to_koalas().toPandas())
# spark -> pandas -> koalas -> spark

df = ks.from_pandas(stats_df.toPandas()).to_spark()
In general, you’ll be working with Spark dataframes when author-

ing code in a PySpark environment. However, it’s useful to be able
to work with different object types as necessary to build model
workflows. Koalas and Pandas UDFs provide powerful tools for
porting workloads to large-scale data ecosystems.
6.3.3 Transforming Data

The PySpark Dataframe API provides a variety of useful functions
for aggregating, filtering, pivoting, and summarizing data. While
some of this functionality maps well to Pandas operations, my
recommendation for quickly getting up and running with munging
data in PySpark is to use the SQL interface to dataframes in Spark
called Spark SQL. If you’re already using the pandasql or framequery
libraries, then Spark SQL should provide a familiar interface. If
you’re new to these libraries, then the SQL interface still provides
an approachable way of working with the Spark ecosystem. We’ll
cover the Dataframe API later in this section, but first start with
the SQL interface to get up and running.
Exploratory data analysis (EDA) is one of the key steps in a data
science workflow for understanding the shape of a data set. To
work through this process in PySpark, we’ll load the stats data set
into a dataframe, expose it as a view, and then calculate summary
statistics. The snippet below shows how to load the NHL stats data
FIGURE 6.4: Summarizing player activity.
set, expose it as a view to Spark, and then run a query against the
dataframe. The aggregated dataframe is then visualized using the
display command in Databricks.
stats_df = spark.read.csv("s3://dsp-ch6/csv/game_skater_stats.csv",
stats_df.createOrReplaceTempView("stats")
new_df = spark.sql("""
select player_id, sum(1) as games, sum(goals) as goals
from stats
group by 1
order by 3 desc
limit 5
""")
display(new_df)
An output of this code block is shown in Figure 6.4. It shows

the highest scoring players in the NHL dataset by ranking the
results based on the total number of goals. One of the powerful
features of Spark is that the SQL query will not operate against the
dataframe until a result set is needed. This means that commands
in a notebook can set up multiple data transformation steps for
Spark dataframes, which are not performed until a later step needs
to execute the graph of operations defined in a code block.
Spark SQL is expressive, fast, and my go-to method for working
with big data sets in a Spark environment. While prior Spark ver-
sions performed better with the Dataframe API versus Spark SQL,
the difference in performance is now trivial and you should use the
transformation tools that provide the best iteration speed for work-
ing with large data sets. With Spark SQL, you can join dataframes,
run nested queries, set up temp tables, and mix expressive Spark
operations with SQL operations. For example, if you want to look
at the distribution of goals versus shots in the NHL stats data, you
can run the following command on the dataframe.
display(spark.sql("""
select cast(goals/shots * 50 as int)/50.0 as Goals_per_shot
,sum(1) as Players
from (
select player_id, sum(shots) as shots, sum(goals) as goals
from stats
group by 1
having goals >= 5
)
group by 1
order by 1
"""))
This query restricts the ratio of goals to shots to players with more
than 5 goals, to prevent outliers such as goalies scoring during
power plays. We’ll use the display command to output the result
set as a table and then use Databricks to display the output as a
chart. Many Spark ecosystems have ways of visualizing results, and
the Databricks environment provides this capability through the
display command, which works well with both tabular and pivot
table data. After running the above command, you can click on
the chart icon and choose dimensions and measures which show
the distribution of goals versus shots, as visualized in Figure 6.5.
While I’m an advocate of using SQL to transform data, since it
scales to different programming environments, it’s useful to get
familiar with some of the basic dataframe operations in PySpark.
The code snippet below shows how to perform common operations
FIGURE 6.5: Distribution of goals per shot.
including dropping columns, selecting a subset of columns, and

adding new columns to a Spark dataframe. Like prior commands,
all of these operations are lazily performed. There are some syn-
tax differences from Pandas, but the general commands used for
transforming data sets should be familiar.
from pyspark.sql.functions import lit
# dropping columns
copy_df = stats_df.drop('game_id', 'player_id')
# selection columns
copy_df = copy_df.select('assists', 'goals', 'shots')
# adding columns
copy_df = copy_df.withColumn("league", lit('NHL'))
display(copy_df)
One of the common operations to perform between dataframes is a

join. This is easy to express in a Spark SQL query, but sometimes
it is preferable to do this programmatically with the Dataframe
API. The code snippet below shows how to join two dataframes
together, when joining on the game_id and player_id fields. The
league column which is a literal will be joined with the rest of the
stats dataframe. This is a trivial example where we are adding a
FIGURE 6.6: The dataframe resulting from the join.
new column onto a small dataframe, but the join operation from
the Dataframe API can scale to massive data sets.
copy_df = stats_df.select('game_id', 'player_id').

withColumn("league", lit('NHL'))
df = copy_df.join(stats_df, ['game_id', 'player_id'])
display(df)
The result set from the join operation above is shown in Figure
6.6. Spark supports a variety of different join types, and in this
example we used an inner join to append the league column to the
players stats dataframe.
It’s also possible to perform aggregation operations on a dataframe,
such as calculating sums and averages of columns. An example of
computing the average time on ice for players in the stats data set,
and total number of goals scored is shown in the snippet below.
The groupBy command uses the player_id as the column for collaps-
ing the data set, and the agg command specifies the aggregations
to perform.
summary_df = stats_df.groupBy("player_id").agg(
{'timeOnIce':'avg', 'goals':'sum'})
display(summary_df)
The snippet creates a dataframe with player_id, timeOnIce, and

goals columns. We’ll again use the plotting functionality in
Databricks to visualize the results, but this time select the scatter
FIGURE 6.7: Time on ice and goal scoring plots.
plot option. The resulting plot of goals versus time on ice is shown
in Figure 6.7.
We’ve worked through introductory examples to get up and run-
ning with dataframes in PySpark, focusing on operations that are
useful for munging data prior to training machine learning models.
These types of operations, in combination with reading and writing
dataframes provide a useful set of skills for performing exploratory
analysis on massive data sets.
6.3.4 Pandas UDFs

While PySpark provides a great deal of functionality for work-
ing with dataframes, it often lacks core functionality provided
in Python libraries, such as the curve-fitting functions in SciPy.
While it’s possible to use the toPandas function to convert

dataframes into the Pandas format for Python libraries, this ap-
proach breaks when using large data sets. Pandas UDFs are a
newer feature in PySpark that help data scientists work around
this problem, by distributing the Pandas conversion across the
worker nodes in a Spark cluster. With a Pandas UDF, you define
a group by operation to partition the data set into dataframes
that are small enough to fit into the memory of worker nodes, and
then author a function that takes in a Pandas dataframe as the
input parameter and returns a transformed Pandas dataframe as
the result. Behind the scenes, PySpark uses the PyArrow library
to efficiently translate dataframes from Spark to Pandas and back
from Pandas to Spark. This approach enables Python libraries,
such as Keras, to be scaled up to a large cluster of machines.
This section will walk through an example problem where we need
to use an existing Python library and show how to translate the
workflow into a scalable solution using Pandas UDFs. The question
we are looking to answer is understanding if there is a positive or
negative relationship between the shots and hits attributes in the
stats data set. To calculate this relationship, we can use the leastsq
function in SciPy, as shown in the snippet below. This example
creates a Pandas dataframe for a single player_id, and then fits
a simple linear regression between these attributes. The output
is the coefficients used to fit the least-squares operation, and in
this case the number of shots was not strongly correlated with the
number of hits.
sample_pd = spark.sql("""
select * from stats
where player_id = 8471214
""").toPandas()
# Import Python libraries

from scipy.optimize import leastsq
import numpy as np
# Define a function to fit

def fit(params, x, y):
return (y - (params[0] + x * params[1] ))
# Fit the curve and show the results

result = leastsq(fit, [1,0], args=(sample_pd.shots,sample_pd.hits))
print(result)
Now we want to perform this operation for every player in the stats
data set. To scale to this volume, we’ll first partition by player_id,
as shown by the groupBy operation in the code snippet below. Next,
We’ll run the analyze_player function for each of these partitioned
data sets using the apply command. While the stats_df dataframe
used as input to this operation and the players_df dataframe re-
turned are Spark dataframes, the sampled_pd dataframe and the
dataframe returned by the analyze player function are Pandas. The
Pandas UDF annotation provides a hint to PySpark for how to dis-
tribute this workload so that it can scale the operation across the
cluster of worker nodes rather than eagerly pulling all of the data
to the driver node. Like most Spark operations, Pandas UDFs are
lazily evaluated and will not be executed until an output value is
needed.
Our initial example now translated to use Pandas UDFs is shown
below. After defining additional modules to include, we specify the
schema of the dataframe that will be returned from the operation.
The schema object defines the structure of the Spark dataframe
that will be returned from applying the analyze player function.
The next step in the code block lists an annotation that defines
this function as a grouped map operation, which means that it
works on dataframes rather than scalar values. As before, we’ll
use the leastsq function to fit the shots and hits attributes. After
calculating the coefficients for this curve fitting, we create a new
Pandas dataframe with the player id, and regression coefficients.
The display command at the end of this code block will force the
Pandas UDF to execute, which will create a partition for each of
the players in the data set, apply the least squares operation, and
merge the results back together into a Spark dataframe.
# Load necessary libraries

from pyspark.sql.functions import pandas_udf, PandasUDFType
from pyspark.sql.types import *
import pandas as pd
# Create the schema for the resulting dataframe

schema = StructType([StructField('ID', LongType(), True),
StructField('p0', DoubleType(), True),
StructField('p1', DoubleType(), True)])
# Define the UDF, input and outputs are Pandas DFs

@pandas_udf(schema, PandasUDFType.GROUPED_MAP)
def analize_player(sample_pd):
# return empty params in not enough data

if (len(sample_pd.shots) <= 1):
return pd.DataFrame({'ID': [sample_pd.player_id[0]],
'p0': [ 0 ], 'p1': [ 0 ]})
# Perform curve fitting

result = leastsq(fit, [1, 0], args=(sample_pd.shots,
sample_pd.hits))
# Return the parameters as a Pandas DF

return pd.DataFrame({'ID': [sample_pd.player_id[0]],
'p0': [result[0][0]], 'p1': [result[0][1]]})
# perform the UDF and show the results

player_df = stats_df.groupby('player_id').apply(analyze_player)
display(player_df)
The key capability that Pandas UDFs provide is that they enable
Python libraries to be used in a distributed environment, as long
as you have a good way of partitioning your data. This means that
FIGURE 6.8: The output dataframe from the Pandas UDF.
libraries such as Featuretools, which were not initially designed to

work in a distributed environment, can be scaled to a large cluster.
The result of applying the Pandas UDF from above on the stats
data set is shown in Figure 6.8. This feature enables a mostly
seamless translation between different dataframe formats.
To further demonstrate the value of Pandas UDFs, we’ll apply
them to distributing a feature generation pipeline and a deep learn-
ing pipeline. However, there are some issues when using Pandas
UDFs in workflows, because they can make debugging more of a
challenge and sometimes fail due to data type mismatches between
Spark and Pandas.
6.3.5 Best Practices

While PySpark provides a familiar environment for Python pro-
grammers, it’s good to follow a few best practices to make sure
you are using Spark efficiently. Here are a set of recommendations
I’ve compiled based on my experience porting a few projects from
Python to PySpark:
• Avoid dictionaries: Using Python data types such as dictionar-
ies means that the code might not be executable in a distributed
mode. Instead of using keys to index values in a dictionary, con-
sider adding another column to a dataframe that can be used
as a filter. This recommendation applies to other Python types
including lists that are not distributable in PySpark.
• Limit Pandas usage: Calling toPandas will cause all data to be
loaded into memory on the driver node, and prevents operations
from being performed in a distributed mode. It’s fine to use this
6.4 MLlib Batch Pipeline 155
function when data has already been aggregated and you want
to make use of familiar Python plotting tools, but it should not
be used for large dataframes.
• Avoid loops: Instead of using for loops, it’s often possible to
use functional approaches such as group by and apply to achieve
the same result. Using this pattern means that code can be par-
allelized by supported execution environments. I’ve noticed that
focusing on using this pattern in Python has also resulted in
cleaner code that is easier to translate to PySpark.
• Minimize eager operations: In order for your pipeline to be as
scalable as possible, it’s good to avoid eager operations that pull
full dataframes into memory. For example, reading in CSVs is an
eager operation, and my work around is to stage the dataframe
to S3 as Parquet before using it in later pipeline steps.
• Use SQL: There are libraries that provide SQL operations
against dataframes in both Python and PySpark. If you’re work-
ing with someone else’s Python code, it can be tricky to decipher
what some of the Pandas operations are achieving. If you plan
on porting your code from Python to PySpark, then using a SQL
library for Pandas can make this translation easier.
By following these best practices when writing PySpark code, I’ve
been able to improve both my Python and PySpark data science
workflows.
6.4 MLlib Batch Pipeline

Now that we’ve covered loading and transforming data with PyS-
park, we can now use the machine learning libraries in PySpark to
build a predictive model. The core library for building predictive
models in PySpark is called MLlib. This library provides a suite of
supervised and unsupervised algorithms. While this library does
not have complete coverage of all of the algorithms in sklearn, it
provides functionality for the majority of the types of operations
needed for data science workflows. In this chapter, we’ll show how
to apply MLlib to a classification problem and save the outputs
from the model application to a data lake.
games_df = spark.read.csv("s3://dsp-ch6/csv/games-expand.csv",
games_df.createOrReplaceTempView("games_df")
games_df = spark.sql("""
select *, row_number() over (order by rand()) as user_id
,case when rand() > 0.7 then 1 else 0 end as test
from games_df
""")
The first step in the pipeline is loading the data set that we want
to use for model training. The snippet above shows how to load the
games data set, and append two additional attributes to the loaded
dataframe using Spark SQL. The result of running this query is
that about 30% of users will be assigned a test value which we’ll
use for model application, and each record is assigned a unique
user ID which we’ll use when saving the model predictions.
The next step is splitting up the data set into train and test
dataframes. For this pipeline, we’ll use the test dataframe as the
data set for model application, where we predict user behavior.
An example of splitting up the dataframes using the test column
is shown in the snippet below. This should result in roughly 16.1k
training users and 6.8k test users.
trainDF = games_df.filter("test == 0")

testDF = games_df.filter("test == 1")
print("Train " + str(trainDF.count()))
print("Test " + str(testDF.count()))
6.4.1 Vector Columns

MLlib requires that the input data is formatted using vector data
types in Spark. To transform our dataframe into this format, we
can use the VectorAssembler class to combine a range of columns into
a single vector column. The code snippet below shows how to use
this class to merge the first 10 columns in the dataframe into a new
vector column called features. After applying this command to the
training dataframe using the transform function, we use the select
function to retrieve only the values we need from the dataframe
for model training and application. For the training dataframe, we
only need the label and features, and with the test dataframe we
also select the user ID.
from pyspark.ml.feature import VectorAssembler
# create a vector representation

assembler = VectorAssembler(
inputCols= trainDF.schema.names[0:10],
outputCol="features" )
trainVec = assembler.transform(trainDF).select('label', 'features')

testVec = assembler.transform(testDF).select(
'label', 'features', 'user_id')
display(testVec)
The display command shows the result of transforming our test

dataset into vector types usable by MLlib. The output dataframe
is visualized in Figure 6.9.
6.4.2 Model Application

Now that we have prepared our training and test data sets, we can
use the logistic regression algorithm provided by MLlib to fit the
training dataframe. We first create a logistic regression object and
define the columns to use as labels and features. Next, we use the
fit function to train the model on the training data set. In the last
FIGURE 6.9: The features in the sparse vector format.
step in the snippet below, we use the transform function to apply

the model to our test data set.
from pyspark.ml.classification import LogisticRegression
# specify the columns for the model

lr = LogisticRegression(featuresCol='features', labelCol='label')
# fit on training data

model = lr.fit(trainVec)
# predict on test data

predDF = model.transform(testVec)
The resulting dataframe now has a probability column, as shown in

Figure 6.10. This column is a 2-element array with the probabilities
for class 0 and 1. To test the accuracy of the logistic regression
model on the test data set, we can use the binary classification
evaluator in Mlib to calculate the ROC metric, as shown in the
snippet below. For my run of the model, the ROC metric has a
value of 0.761.
from pyspark.ml.evaluation import BinaryClassificationEvaluator
roc = BinaryClassificationEvaluator().evaluate(predDF)
print(roc)
FIGURE 6.10: The dataframe with propensity scores.
In a production pipeline, there will not be labels for the users

that need predictions, meaning that you’ll need to perform cross
validation to select the best model for making predictions. An
example of this approach is covered in Section 6.7. In this case,
we are using a single data set to keep code examples short, but a
similar pipeline can be used in a production workflows.
Now that we have the model predictions for our test users, we
need to retrieve the predicted label in order to create a dataframe
to persist to a data lake. Since the probability column created by
MLlib is an array, we’ll need to define a UDF that retrieves the
second element as our propensity column, as shown in the PySpark
snippet below.
from pyspark.sql.functions import udf

from pyspark.sql.types import FloatType
# split out the array into a column

secondElement = udf(lambda v:float(v[1]),FloatType())
predDF = predDF.select("*",
secondElement("probability").alias("propensity"))
display(predDF)
After running this code block, the dataframe will have an addi-
tional column called propensity as shown in Figure 6.10. The final
step in this batch prediction pipeline is to save the results to S3.
We’ll use the select function to retrieve the relevant columns from
the predictions dataframe, and then use the write function on the
dataframe to persist the results as Parquet on S3.
# save results to S3
results_df = predDF.select("user_id", "propensity")
results_path = "s3a://dsp-ch6/game-predictions/"
results_df.write.mode('overwrite').parquet(results_path)
We now have all of the building blocks needed to create a PySpark

pipeline that can fetch data from a storage layer, train a predictive
model, and write the results to persistent storage. We’ll cover how
to schedule this type of pipeline in Section 6.8.
When developing models, it’s useful to inspect the output to see
if the distribution of model predictions matches expectations. We
can use Spark SQL to perform an aggregation on the model out-
puts and then use the display command to perform this process
directly in Databricks, as shown in the snippet below. The result
of performing these steps on the model predictions is shown in
Figure 6.11.
# plot the predictions

predDF .createOrReplaceTempView("predDF ")
plotDF = spark.sql("""
select cast(propensity*100 as int)/100 as propensity,
label, sum(1) as users
from predDF
group by 1, 2
order by 1, 2
""")
# table output
display(plotDF)
MLlib can be applied to a wide variety of problems using a large

suite of algorithms. While we explored logistic regression in this
section, the libraries provides a number of different classification
6.5 Distributed Deep Learning 161
FIGURE 6.11: The distribution of propensity scores.
approaches, and there are other types of operations supported in-

cluding regression and clustering.
6.5 Distributed Deep Learning

While MLlib provides scalable implementations for classic machine
learning algorithms, it does not natively support deep learning
libraries such as TensorFlow and PyTorch. There are libraries that
parallelize the training of deep learning models on Spark, but the
data set needs to be able to fit in memory on each worker node,
and these approaches are best used for distributed hyperparameter
tuning on medium-sized data sets.
For the model application stage, where we already have a deep
learning model trained but need to apply the resulting model to
a large user base, we can use Pandas UDFs. With Pandas UDFs,
we can partition and distribute our data set, run the resulting
dataframes against a Keras model, and then compile the results
back into a single large Spark dataframe. This section will show
how we can take the Keras model that we built in Section 1.6.3,
and scale it to larger data sets using PySpark and Pandas UDFs.
However, we still have the requirement that the data used for train-
ing the model can fit into memory on the driver node.
We’ll use the same data sets from the prior section, where we split
the games data set into training and test sets of users. This is a
relatively small data set, so we can use the toPandas operation to

load the dataframe onto the driver node, as shown in the snippet
below. The result is a dataframe and list that we can provide as
input to train a Keras deep learning model.
# build model on the driver node

train_pd = trainDF.toPandas()
x_train = train_pd.iloc[:,0:10]
y_train = train_pd['label']
When using PyPI to install TensorFlow on the Spark cluster, the

installed version of the library should be 2.0 or greater. This differs
from Version 1 of TensorFlow, which we used in prior chapters.
The main impact in terms of the code snippets is that TensorFlow
2 now has a built-in AUC function that no longer requires the
workflow we previously applied.
6.5.1 Model Training

We’ll use the same approach as before to train a Keras model. The
code snippet below shows how to set up a network with an in-
put layer, dropout later, single hidden layer, and an output layer,
optimized with rmsprop and a cross entropy loss. In the model ap-
plication phase, we’ll reuse the model object in a Pandas UDFs to
distribute the workload.
import keras
model.add(layers.Dense(64, activation='relu', input_shape=(10,)))
model.compile(optimizer='rmsprop', loss='binary_crossentropy')
history = model.fit(x_train, y_train, epochs=100, batch_size=100,

To test for overfitting, we can plot the results of the training and
validation data sets, as shown in Figure 6.12. The snippet below
shows how to use matplotlib to display the losses over time for
these data sets. While the training loss continued to decrease over
additional epochs, the validation loss stopped improving after 20
epochs, but did not noticeably increase over time.
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(1, len(loss) + 1)
fig = plt.figure(figsize=(10,6) )
plt.plot(epochs, loss, 'bo', label='Training Loss')
plt.plot(epochs, val_loss, 'b', label='Validation Loss')
plt.legend()
plt.show()
display(fig)
6.5.2 Model Application

Now that we have a trained deep learning model, we can use PyS-
park to apply it in a scalable pipeline. The first step is determining
how to partition the set of users that need to be scored. For this
data set, we can split the user base into 100 different buckets, as
shown in the snippet below. This randomly assigns each user into 1
of 100 buckets, which means that after applying the group by step,
each dataframe that gets translated to Pandas will be roughly 1%
of the size of the original dataframe. If you have a large data set,
you may need to use thousands of buckets to distribute the data
set, and maybe more.
FIGURE 6.12: Training a Keras model on a subset of data.
# set up partitioning for the train dataframe

testDF.createOrReplaceTempView("testDF ")
partitionedDF = spark.sql("""
select *, cast(rand()*100 as int) as partition_id
from testDF
""")
The next step is to define the Pandas UDF that will apply the
Keras model. We’ll define an output schema of a user ID and
propensity score, as shown below. The UDF uses the predict func-
tion on the model object we previously trained to create a pre-
diction column on the passed in dataframe. The return command
selects the two relevant columns that we defined for the schema
object. The group by command partitions the data set using our
bucketing approach, and the apply command performs the Keras
model application across the cluster of worker nodes. The result is
a Spark dataframe visualized with the display command, as shown
in Figure 6.13.
FIGURE 6.13: The resulting dataframe for distributed Keras.

from pyspark.sql.types import *
schema = StructType([StructField('user_id', LongType(), True),

StructField('propensity', DoubleType(),True)])
def apply_keras(pd):
pd['propensity'] = model.predict(pd.iloc[:,0:10])
return pd[['user_id', 'propensity']]
results_df=partitionedDF.groupby('partition_id').apply(apply_keras)
display(results_df)
One thing to note is that there are limitations on the types of

objects that you can reference in a Pandas UDFs. In this example,
we referenced the model object, which was created on the driver
node when training the model. When variables in PySpark are
transferred from the driver node to workers nodes for distributed
operations, a copy of the variable is made, because synchronizing
variables across a cluster would be inefficient. This means that any
changes made to a variable within a Pandas UDF will not apply
to the original object. It’s also why data types such as Python lists
and dictionaries should be avoided when using UDFs. Functions
work in a similar way, and in Section 6.3.4 we used the fit function
in a Pandas UDF where the function was initially defined on the

driver node. Spark also provides broadcast variables for sharing
variables in a cluster, but ideally distributed code segments should
avoid sharing state through variables if possible.
6.6 Distributed Feature Engineering

Feature engineering is a key step in a data science workflow, and
sometimes it is necessary to use Python libraries to implement
this functionality. For example, the AutoModel system at Zynga
uses the Featuretools library to generate hundreds of features from
raw tracking events, which are then used as input to classification
models. To scale up the automated feature engineering approach
that we first explored in Section 1.7, we can use Pandas UDFs to
distribute the feature application process. Like the prior section,
we need to sample data when determining the transformation to
perform, but when applying the transformation we can scale to
massive data sets.
For this section, we’ll use the game plays data set from the NHL
Kaggle example, which includes detailed play-by-play descriptions
of the events that occurred during each match. Our goal is to
transform the deep and narrow dataframe into a shallow and wide
dataframe that summarizes each game as a single record with hun-
dreds of columns. An example of loading this data in PySpark and
selecting the relevant columns is shown in the snippet below. Be-
fore calling toPandas, we use the filter function to sample 0.3% of
the records, and then cast the result to a Pandas frame, which has
a shape of 10,717 rows and 16 columns.
plays_df = spark.read.csv("s3://dsp-ch6/csv/game_plays.csv",
header=True, inferSchema = True).drop(
'secondaryType', 'periodType', 'dateTime', 'rink_side')
plays_pd = plays_df.filter("rand() < 0.003").toPandas()
plays_pd.shape
6.6 Distributed Feature Engineering 167
6.6.1 Feature Generation

We’ll use the same two-step process covered in Section 1.7 where
we first one-hot encode the categorical features in the dataframe,
and then apply deep feature synthesis to the data set. The code
snippet below shows how to perform the encoding process using
the Featuretools library. The output is a transformation of the
initial dataframe that now has 20 dummy variables instead of the
event and description variables.
import featuretools as ft
from featuretools import Feature
es = es.entity_from_dataframe(entity_id="plays",dataframe=plays_pd,
index="play_id", variable_types = {
f1 = Feature(es["plays"]["event"])
f2 = Feature(es["plays"]["description"])
encoded, defs = ft.encode_features(plays_pd, [f1, f2], top_n=10)

encoded.reset_index(inplace=True)
The next step is using the dfs function to perform deep feature
synthesis on our encoded dataframe. The input dataframe will have
a record per play, while the output dataframe will have a single
record per game after collapsing the detailed events into a wide
column representation using a variety of different aggregations.
FIGURE 6.14: The schema for the generated features.
features, transform=ft.dfs(entityset=es,
target_entity="games",max_depth=2)
features.reset_index(inplace=True)
One of the new steps that we need to perform versus the prior
approach, is that we need to determine what the schema will be
for the generated features, since this is needed as an input to the
Pandas UDF annotation. To figure out what the generated schema
is for the generated dataframe, we can create a Spark dataframe
and then retrieve the schema from the dataframe. Before convert-
ing the Pandas dataframe, we need to modify the column names
in the generated dataframe to remove special characters, as shown
in the snippet below. The resulting Spark schema for the feature
application step is displayed in Figure 6.14.
features.columns = features.columns.str.replace("[(). =]", "")

schema = sqlContext.createDataFrame(features).schema
features.columns
We now have the required schema for defining a Pandas UDF.

Unlike the past UDFs we defined, the schema may change between
different runs based on the feature transformation aggregations
selected by Featuretools. In these steps, we also created a defs
object that defines the feature transformations to use for encoding
and a transform object that defines the transformations to perform
6.6 Distributed Feature Engineering 169
deep feature synthesis. Like the model object in the past section,
copies of these objects will be passed to the Pandas UDF executing
on worker nodes.
6.6.2 Feature Application

To enable our approach to scale across a cluster of worker nodes,
we need to define a column to use for partitioning. Like the prior
section, we can bucket events into different sets of data to ensure
that the UDF process can scale. One difference from before is that
we need all of the plays from a specific game to be grouped into
the same partition. To achieve this result, we can partition by the
game_id rather than the player_id. An example of this approach
in shown in the code snippet below. Additionally, we can use the
hash function on the game ID to randomize the value, resulting in
more balanced bucket sizes.
# bucket IDs
plays_df.createOrReplaceTempView("plays_df")
plays_df = spark.sql("""
select *, abs(hash(game_id))%1000 as partition_id
from plays_df
""")
We can now apply feature transformation to the full data set, using
the Pandas UDF defined below. The plays dataframe is partitioned
by the bucket before being passed to the generate features func-
tion. This function uses the previously generated feature transfor-
mations to ensure that the same transformation is applied across
all of the worker nodes. The input Pandas dataframe is a narrow
and deep representation of play data, while the returned dataframe
is a shallow and wide representation of game summaries.
def gen_features(plays_pd):
FIGURE 6.15: Generated features in the Spark dataframe.
dataframe=plays_pd, index="play_id", variable_types = {
encoded_features = ft.calculate_feature_matrix(defs, es)
encoded_features.reset_index(inplace=True)
generated = ft.calculate_feature_matrix(transform,es).fillna(0)
generated.reset_index(inplace=True)
generated.columns = generated.columns.str.replace("[(). =]","")
return generated
features_df = plays_df.groupby('partition_id').apply(gen_features)
display(features_df)
The output of the display command is shown in Figure 6.15. We’ve

now worked through feature generation and deep learning in scal-
able model pipelines. Now that we have a transformed data set, we
can join the result with additional features, such as the label that
we are looking to predict, and develop a complete model pipeline.
6.7 GCP Model Pipeline 171
6.7 GCP Model Pipeline

A common workflow for batch model pipelines is reading input
data from a lake, applying a machine learning model, and then
writing the results to an application database. In GCP, BigQuery
serves as the data lake and Cloud Datastore can serve as an ap-
plication database. We’ll build and end-to-end pipeline with these
components in the next chapter, but for now we’ll get hands on
with a subset of the GCP components directly in Spark.
While there is a Spark connector for BigQuery3 , enabling large-
scale PySpark pipelines to be built using BigQuery directly, there
are some issues with this library that make it quite complicated
to set up for our Databricks environment. For example, we would
need to rebuild some of the JAR files and shade the dependencies.
One alternative is to use the Python BigQuey connector that we
explored in Section 5.1, but this approach is not distributed and
will eagerly pull the query results to the driver node as a Pandas
dataframe. For this chapter, we’ll explore a workflow where we
unload query results to Cloud Storage, and then read in the data
set from GCS as the initial step in the pipeline. Similarly, for model
output we’ll save the results to GCS, where the output is available
for pushing to Cloud Datastore. To productize this type of model
workflow, Airflow could be used to chain these different actions
together.
6.7.1 BigQuery Export

The first step we’ll perform is exporting the results of a BigQuery
query to GCS, which can be performed manually using the Big-
Query UI. This is possible to perform directly in Spark, but as I
mentioned the setup is quite involved to configure with the current
version of the connector library. We’ll use the natality data set for
this pipeline, which lists attributes about child deliveries, such as
birth weight.
3
https://github.com/spotify/spark-bigquery/
create table dsp_demo.natality as (

select *
from `bigquery-public-data.samples.natality`
order by rand()
limit 10000
)
To create a data set, we’ll sample 10k records from the natality
public data set in BigQuery. To export this result set to GCS, we
need to create a table on BigQuery with the data that we want
to export. The SQL for creating this data sample is shown in the
snippet above. To export this data to GCS, perform the following
steps:
1. Browse to the GCP Console

2. Search for “BigQuery”
3. Paste the Query from the snippet above into the editor
4. Click Run
5. In the left pane, select the table, “dsp_demo.natality”
6. Click “Export”, and then “Export to GCS”
7. Set the location, “/dsp_model_store/natality/avro”
8. Use “Avro” as export format
9. Click “Export”
After performing these steps, the sampled natality data will be

saved to GCS in Avro format. The confirmation dialog from ex-
porting the data set is shown in Figure 6.16. We now have the
data saved to GCS in a format that works well with Spark.
6.7.2 GCP Credentials

We now have a data set that we can use as input to a PySpark
pipeline, but we don’t yet have access to the bucket on GCS from
our Spark environment. With AWS, we were able to set up pro-
grammatic access to S3 using an access and secret key. With GCP,
the process is a bit more complicated because we need to move the
json credentials file to the driver node of the cluster in order to
FIGURE 6.16: Confirming the Avro export on GCS.
read and write files on GCS. One of the challenges with using
Spark is that you may not have SSH access to the driver node,
which means that we’ll need to use persistent storage to move the
file to the driver machine. This isn’t recommended for production
environments, but instead is being shown as a proof of concept.
The best practice for managing credentials in a production envi-
ronment is to use IAM roles.
aws s3 cp dsdemo.json s3://dsp-ch6/secrets/dsdemo.json
aws s3 ls s3://dsp-ch6/secrets/
To move the json file to the driver node, we can first copy the
credentials file to S3, as shown in the snippet above. Now we can
switch back to Databricks and author the model pipeline. To copy
the file to the driver node, we can read in the file using the sc
Spark context to read the file line by line. This is different from
all of our prior operations where we have read in data sets as
dataframes. After reading the file, we then create a file on the
driver node using the Python open and write functions. Again, this
is an unusual action to perform in Spark, because you typically
want to write to persistent storage rather than local storage. The
result of performing these steps is that the credentials file will now
be available locally on the driver node in the cluster.
creds_file = '/databricks/creds.json'
creds = sc.textFile('s3://dsp-ch6/secrets/dsdemo.json')
with open(creds_file, 'w') as file:

for line in creds.take(100):
file.write(line + "\n")
Now that we have the json credentials file moved to the driver local
storage, we can set up the Hadoop configuration needed to access
data on GCS. The code snippet below shows how to configure
the project ID, file system implementation, and credentials file
location. After running these commands, we now have access to
read and write files on GCS.
sc._jsc.hadoopConfiguration().set("fs.gs.impl",
"com.google.cloud.hadoop.fs.gcs.GoogleHadoopFileSystem")
sc._jsc.hadoopConfiguration().set("fs.gs.project.id",
"your_project_id")
"mapred.bq.auth.service.account.json.keyfile", creds_file)
"fs.gs.auth.service.account.json.keyfile", creds_file)
6.7.3 Model Pipeline

To read in the natality data set, we can use the read function with
the Avro setting to fetch the data set. Since we are using the Avro
format, the dataframe will be lazily loaded and the data is not
retrieved until the display command is used to sample the data
set, as shown in the snippet below.
natality_path = "gs://dsp_model_store/natality/avro"
natality_df = spark.read.format("avro").load(natality_path)
display(natality_df)
Before we can use MLlib to build a regression model, we need to

perform a few transformations on the data set to select a subset
of the features, cast data types, and split records into training and
test groups. We’ll also use the fillna function as shown below in
order to replace any null values in the dataframe with zeros. For
this modeling exercise, we’ll build a regression model that predicts
the birth weight of a baby using a few different features including
the marriage status of the mother and parent ages. The prepared
dataframe is shown in Figure 6.17.
natality_df.createOrReplaceTempView("natality_df")
natality_df = spark.sql("""
SELECT year, plurality, apgar_5min,
mother_age, father_age,
gestation_weeks, ever_born
,case when mother_married = true
then 1 else 0 end as mother_married
,weight_pounds as weight
,case when rand() < 0.5 then 1 else 0 end as test
from natality_df
""").fillna(0)
trainDF = natality_df.filter("test == 0")

testDF = natality_df.filter("test == 1")
display(natality_df)
Next, we’ll translate our dataframe into the vector data types that
MLlib requires as input. The process for transforming the natality
data set is shown in the snippet below. After executing the trans-
form function, we now have training and test data sets we can use
FIGURE 6.17: The prepared Natality dataframe.
as input to a regression model. The label we are building a model

to predict is the weight column.
from pyspark.ml.feature import VectorAssembler
# create a vector representation

assembler = VectorAssembler(inputCols= trainDF.schema.names[0:8],
outputCol="features" )
trainVec = assembler.transform(trainDF).select('weight','features')
testVec = assembler.transform(testDF).select('weight', 'features')
MLlib provides a set of utilities for performing cross validation

and hyperparameter tuning in a model workflow. The code snip-
pet below shows how to perform this process for a random forest
regression model. Instead of calling fit directly on the model ob-
ject, we wrap the model object with a cross validator object that
explores different parameter settings, such as tree depth and num-
ber of trees. This workflow is similar to the grid search functions in
sklearn. After searching through the parameter space, and using
cross validation based on the number of folds, the random forest
model is retrained on the complete training data set before being
applied to make predictions on the test data set. The result is a
dataframe with the actual weight and predicted birth weight.
from pyspark.ml.tuning import ParamGridBuilder

from pyspark.ml.regression import RandomForestRegressor
from pyspark.ml.tuning import CrossValidator
from pyspark.ml.evaluation import RegressionEvaluator
folds = 3
rf_trees = [ 50, 100 ]
rf_depth = [ 4, 5 ]
rf= RandomForestRegressor(featuresCol='features',labelCol='weight')
paramGrid = ParamGridBuilder().addGrid(rf.numTrees, rf_trees).

ddGrid(rf.maxDepth, rf_depth).build()
crossval = CrossValidator(estimator=rf, estimatorParamMaps =
paramGrid, evaluator=RegressionEvaluator(
labelCol='weight'), numFolds = folds)
rfModel = crossval.fit(trainVec)
predsDF = rfModel.transform(testVec).select("weight", "prediction")
In the final step of our GCP model pipeline, we’ll save the results
to GCS, so that other applications or processes in a workflow can
make use of the predictions. The code snippet below shows how
to write the dataframe to GCS in Avro format. To ensure that
different runs of the pipeline do not overwrite past predictions, we
append a timestamp to the export path.
import time
out_path = "gs://dsp_model_store/natality/preds-{time}/".
format(time = int(time.time()*1000))
predsDF.write.mode('overwrite').format("avro").save(out_path)
print(out_path)
Using GCP components with PySpark took a bit of effort to con-

figure, but in this case we are running Spark in a different cloud
provider than where we are reading and writing data. In a pro-
duction environment, you’ll most likely be running Spark in the
same cloud as where you are working with data sets, which means
that you can leverage IAM roles for properly managing access to
different services.
6.8 Productizing PySpark

Once you’ve tested a batch model pipeline in a notebook envi-
ronment, there are a few different ways of scheduling the pipeline
to run on a regular schedule. For example, you may want a churn
prediction model for a mobile game to run every morning and pub-
lish the scores to an application database. Similar to the workflow
tools we covered in Chapter 5, a PySpark pipeline should have
monitoring in place for any failures that may occur. There’s a few
different approaches for scheduling PySpark jobs to run:
• Workflow Tools: Airflow, Azkaban, and Luigi all support run-
ning spark jobs as part of a workflow.
• Cloud Tools: EMR on AWS and Dataproc on GCP support
scheduled Spark jobs.
• Vendor Tools: Databricks supports setting up job schedules
with monitoring through the web UI.
• Spark Submit: If you already have a cluster provisioned, you
can issue spark-submit commands using a tool such as crontab.
Vendor and cloud tools are typically easier to get up and running,
because they provide options for provisioning clusters as part of
the workflow. For example, with Databricks you can define the
type of cluster to spin up for running a notebook on a schedule.
When using a workflow tool, such as Airflow, you’ll need to add
additional steps to your workflow in order to spin up and termi-
nate clusters. Most workflow tools provide connectors to EMR for
managing clusters as part of a workflow. The Spark submit option
is useful when first getting started with scheduling Spark jobs, but
it doesn’t support managing clusters as part of a workflow.
Spark jobs can run on ephemeral or persistent clusters. An
ephemeral cluster is a Spark cluster that is provisioned to per-
form a set of tasks and then terminated, such as running a churn
6.9 Conclusion 179
model pipeline. A persistent cluster is a long-running cluster than

may support interactive notebooks, such as the Databricks cluster
we set up at the start of this chapter. Persistent clusters are useful
for development, but can be expensive if the hardware spun up for
the cluster is under utilized. Some vendors support auto scaling
of clusters to reduce the cost of long-running persistent clusters.
Ephemeral clusters are useful, because spinning up a new cluster
to perform a task enables isolation of failure across tasks, and it
means that different model pipelines can use different library ver-
sions and Spark runtimes.
In addition to setting up tools for scheduling jobs and alerting
on job failures, it’s useful to set up additional data and model
quality checks for Spark model pipelines. For example, I’ve set up
Spark jobs that perform audit tasks, such as making sure that
an application database has predictions for the current day, and
trigger alerts if prediction data is stale. It’s also a good practice to
log metrics, such as the ROC of a cross-validated model, as part
of a Spark pipeline.
6.9 Conclusion
PySpark is a powerful tool for data scientists to build scalable
analyses and model pipelines. It a highly desirable skill set for
companies, because it enables data science teams to own more of
the process of building and owning data products. There’s a variety
of ways to set up an environment for PySpark, and in this chapter
we explored a free notebook environment from one of the popular
Spark vendors.
This chapter focused on batch model pipelines, where the goal is
to create a set of predictions for a large number of users on a
regular schedule. We explored pipelines for both AWS and GCP
deployments, where the data sources and data outputs are data
lakes. One of the issues with these types of pipelines is that predic-
tions may be quite stale by the time that a prediction is used. In
Chapter 8, we’ll explore streaming pipelines for PySpark, where

the latency of model predictions is minimized.
PySpark is a highly expressive language for authoring model
pipelines, because it supports all Python functionality, but does
require some workarounds to get code to execute across a cluster
of workers nodes. In the next chapter we’ll explore Dataflow, a run-
time for the Apache Beam library, which also enables large-scale
distributed Python pipelines, but is more constrained in the types
of operations that you can perform.
7
Cloud Dataflow for Batch Modeling
Dataflow is a tool for building data pipelines that can run locally,
or scale up to large clusters in a managed environment. While
Cloud Dataflow was initially incubated at Google as a GCP specific
tool, it now builds upon the open-source Apache Beam library,
making it usable in other cloud environments. The tool provides
input connectors to different data sources, such as BigQuery and
files on Cloud Storage, operators for transforming and aggregating
data, and output connectors to systems such as Cloud Datastore
and BigQuery.
In this chapter, we’ll build a pipeline with Dataflow that reads in
data from BigQuery, applies a sklearn model to create predictions,
and then writes the predictions to BigQuery and Cloud Datastore.
We’ll start by running the pipeline locally on a subset of data and
then scale up to a larger data set using GCP.
Dataflow is designed to enable highly-scalable data pipelines, such
as performing ETL work where you need to move data between
different systems in your cloud deployment. It’s also been extended
to work well for building ML pipelines, and there’s built-in support
for TensorFlow and other machine learning methods. The result is
that Dataflow enables data scientists to build large scale pipelines
without needing the support of an engineering team to scale things
up for production.
The core component in Dataflow is a pipeline, which defines the op-
erations to perform as part of a workflow. A workflow in Dataflow
is a DAG that includes data sources, data sinks, and data trans-
formations. Here are some of the key components:
181
182 7 Cloud Dataflow for Batch Modeling
• Pipeline: Defines the set of operations to perform as part of a

Dataflow job.
• Collection: The interface between different stages in a workflow.
The input to any step in a workflow is a collection of objects and
the output is a new collection of objects.
• DoFn: An operation to perform on each element in a collection,
resulting in a new collection.
• Transform: An operation to perform on sets of elements in a
collection, such as an aggregation.
Dataflow works with multiple languages, but we’ll focus on the
Python implementation for this book. There are some caveats with
the Python version, because worker nodes may need to compile
libraries from source, but it does provide a good introduction to
the different components in Apache Beam. To create a workflow
with Beam, you use the pipe syntax in Python to chain different
steps together. The result is a DAG of operations to perform that
can be distributed across machines in a cluster.
The two ways of transforming data in a Dataflow pipeline are DoFn
and Transform steps. A DoFn step defines an operation to perform on
each object in a collection. For example, we’ll query the Natality
public data set and the resulting collection will contain dictionary
objects. We’ll define a DoFn operation that uses sklearn to create a
prediction for each of these dictionary objects and output a new
dictionary object. A Transform defines an operation to perform on
a set of objects, such as performing feature generation to aggre-
gate raw tracking events into user-level summaries. These types of
operations are typically used in combination with a partition trans-
form step to divide up a collection of objects into a manageable
size. We won’t explore this process in this book, but a transform
could be used to apply Featuretools to perform automated feature
engineering as part of a Dataflow pipeline.
In this chapter we’ll get hands on with building Dataflow pipelines
that can run locally and in a fully-managed GCP cluster. We’ll
start by building a simple pipeline that works with text data, and
then build a pipeline that applies a sklearn model in a distributed
workflow.
7.1 Apache Beam 183
7.1 Apache Beam

Apache Beam is an open-source library for building data processing
workflows using Java, Python, and Go. Beam workflows can be ex-
ecuted across several execution engines including Spark, Dataflow,
and MapReduce. With Beam, you can test workflows locally us-
ing the Direct Runner for execution, and then deploy the workflow
in GCP using the Dataflow Runner. Beam pipelines can be batch,
where a workflow is executed until it is completed, or streaming,
where the pipeline runs continuously and operations are performed
in near real-time as data is received. We’ll focus on batch pipelines
in this chapter and cover streaming pipelines in the next chapter.
The first thing we’ll need to do in order to get up and running
is install the Apache Beam library. Run the commands shown be-
low from the command line in order to install the library, set up
credentials for GCP, and to run a test pipeline locally. The pip
command includes the gcp annotation to specify that the Dataflow
modules should also be installed. If the last step is successful, the
pipeline will output the word counts for Shakespeare’s King Lear.
# install APache Bean

pip install --user apache-beam[gcp]
# set up GCP credentials

export GOOGLE_APPLICATION_CREDENTIALS=/home/ec2-user/dsdemo.json
# run the word count example

python3 -m apache_beam.examples.wordcount --output outputs
The example pipeline performs a number of different steps in order

to perform this counting logic. First, the pipeline reads in the play
as a collection of string objects, where each line from the play is a
string. Next, the pipeline splits each line into a collection of words,
which are then passed to map and group transforms that count
the occurrence of each word. The map and group operations are
built-in Bean transform operations. The last step is writing the

collection of word counts to the console.
Instead of walking through the example code in detail, we’ll build
our own pipeline that more closely resembles the workflow of build-
ing a batch model application pipeline. The listing below shows the
full code for building and running a pipeline that reads in the play
from Cloud Storage, appends a message to the end of every line of
text, and writes the results back to Cloud Storage. The complete
pipeline can be executed from within a Jupyter notebook, which
is a useful way of getting up and running with simple pipelines
when learning Dataflow.
import apache_beam as beam

import argparse
from apache_beam.options.pipeline_options import PipelineOptions
from apache_beam.io import ReadFromText
from apache_beam.io import WriteToText
# define a function for transforming the data

class AppendDoFn(beam.DoFn):
def process(self, element):
return element + " - Hello World!"
# set up pipeline parameters

parser = argparse.ArgumentParser()
parser.add_argument('--input', dest='input',
default='gs://dataflow-samples/shakespeare/kinglear.txt')
parser.add_argument('--output', dest='output',
default='gs://dsp_model_store/shakespeare/kinglear.txt')
known_args, pipeline_args = parser.parse_known_args(None)
pipeline_options = PipelineOptions(pipeline_args)
# define the pipeline steps

p = beam.Pipeline(options=pipeline_options)
lines = p | 'read' >> ReadFromText(known_args.input)
appended = lines | 'append' >> beam.ParDo(AppendDoFn())
7.1 Apache Beam 185
appended | 'write' >> WriteToText(known_args.output)
# run the pipeline

result = p.run()
result.wait_until_finish()
The first step in this code is to load the necessary modules needed
in order to set up a Beam pipeline. We import IO methods for
reading and writing text files, and utilities for passing parameters
to the Beam pipeline. Next, we define a class that will perform
a DoFn operation on every element passed to the process function.
This class extends the beam.DoFn class, which provides an interface
for processing elements in a collection. The third step is setting up
parameters for the pipeline to use for execution. For this example,
we need to set up the input location for reading the text and output
location for writing the result.
Once we have set up the pipeline options, we can set up the DAG
that defines the sequence of actions to perform. For this example,
we’ll create a simple sequence where the input text is passed to
our append step and the output is passed to the text writer. A
visualization of this pipeline is shown in Figure 7.1. To construct
the DAG, we use pipe (|) commands to chain the different steps
together. Each step in the pipeline is a ParDo or Transform command
that defines the Beam operation to perform. In more complicated
workflows, an operation can have multiple outputs and multiple
inputs.
Once the pipeline is constructed, we can use the run function to
execute the pipeline. When running this example in Jupyter, the
Direct Runner will be used by Beam to execute the pipeline on
the local machine. The last command waits for the pipeline to
complete before proceeding.
With the Direct Runner, all of the global objects defined in the
Python file can be used in the DoFn classes, because the code is run-
ning as a single process. When using a distributed runner, some ad-
ditional steps need to be performed to make sure that the class has
FIGURE 7.1: Running the managed pipeline on GCP.
FIGURE 7.2: The resulting file on Google Storage.
access to the modules needed to perform operations. We’ll cover

this issue in the next section, since the process function in this
example does not use any modules. In general, process functions
should only make use of modules defined in the init function, the
passed in elements, and any side inputs that are provided to the
model. We won’t cover side inputs in this chapter, which provide a
way of passing addition data to DoFn operations, and instead we’ll
load the model from Cloud Storage when the class is instantiated.
7.1 Apache Beam 187
After running the pipeline, the appended text will be available on

Cloud Storage. You can validate that the pipeline was successful
by browsing to the bucket in the GCP console, as shown in Figure
7.2. The result is saved as a single file, but for larger outputs the
result will be split into multiple files, where the best practice is
to use Avro or Parquet formats. The Avro format works well with
Dataflow, because data gets streamed between different steps in
the pipeline, even when running in batch mode, and results can
be written to storage once they are ready on each worker machine.
Unlike Spark, where stages with dependencies are not executed
concurrently, steps in a Dataflow workflow with dependencies can
execute simultaneously.
While it’s possible to run pipelines from Jupyter, which is use-
ful for learning how pipelines work, it’s more common to use a
text editor or IDE to create Python files that are executed via
the command line. The first command below shows how to run
the example pipeline with the Direct Runner, and the second com-
mand shows how to run the pipeline on Cloud Dataflow using the
runner parameter. We also need to include a staging location on
Cloud Storage for Dataflow to manage the job and specify our
GCP project name.
# run locally
python3 append.py \
# run managed
python3 append.py \
--runner DataflowRunner \
--project your_project_name \
--temp_location gs://dsp_model_store/tmp/
By default, Beam pipelines run as a batch process. To run in

streaming mode, you need to pass the streaming flag, which we’ll
cover in the next chapter. The result of running the workflow on
Dataflow is shown in Figure 7.1. You can view the progress of your
workflows on GCP by browsing to the console and navigating to

the Dataflow view, which will show a list of jobs.
We now have hands on experience with building a data pipeline
using Apache Beam, and have run the pipeline locally and in a
managed cloud environment. The next step is to use the process
function to apply an ML model to the passed in data set.
7.2 Batch Model Pipeline

Cloud Dataflow provides a useful framework for scaling up sklearn
models to massive data sets. Instead of fitting all input data into
a dataframe, we can score each record individually in the process
function, and use Apache Beam to stream these outputs to a data
sink, such as BigQuery. As long as we have a way of distributing
our model across the worker nodes, we can use Dataflow to perform
distributed model application. This can be achieved by passing
model objects as side inputs to operators or by reading the model
from persistent storage such as Cloud Storage. In this section we’ll
first train a linear regression model using a Jupyter environment,
and then store the results to Cloud Storage so that we can run the
model on a large data set and save the predictions to BigQuery
and Cloud Datastore.
7.2.1 Model Training

The modeling task that we’ll be performing is predicting the birth
weight of a child given a number of factors, using the Natality
public data set. To build a model with sklearn, we can sample
the data set before loading it into a Pandas dataframe and fitting
the model. The code snippet below shows how to sample the data
set from a Jupyter notebook and visualize a subset of records, as
shown in Figure 7.3.
7.2 Batch Model Pipeline 189
FIGURE 7.3: The sampled Natality data set for training.

sql = """
order by rand()
limit 10000
"""
natalityDF = client.query(sql).to_dataframe().fillna(0)
natalityDF.head()
Once we have the data to train on, we can use the LinearRegression
class in sklearn to fit a model. We’ll use the full dataframe for
fitting, because the holdout data is the rest of the data set that
was not sampled. Once trained, we can use pickle to serialize the
model and save it to disk. The last step is to move the model file
from local storage to cloud storage, as shown below. We now have
a model trained that can be used as part of a distributed model
application workflow.
from sklearn.linear_model import LinearRegression

import pickle
# fit and pickle a model

model = LinearRegression()
model.fit(natalityDF.iloc[:,1:8], natalityDF['weight'])
pickle.dump(model, open("natality.pkl", 'wb'))
# Save to GCS
bucket = storage.Client().get_bucket('dsp_model_store')
blob = bucket.blob('natality/sklearn-linear')
blob.upload_from_filename('natality.pkl')
7.2.2 BigQuery Publish

We’ll start by building a Beam pipeline that reads in data from
BigQuery, applies a model, and then writes the results to BigQuery.
In the next section, we’ll add Cloud Datastore as an additional
data sink for the pipeline. This pipeline will be a bit more complex
than the prior example, because we need to use multiple Python
modules in the process function, which requires a bit more setup.
We’ll walk through different parts of the pipeline this time, to pro-
vide additional details about each step. The first task is to define
the libraries needed to build and execute the pipeline. We are also
importing the json module, because we need this to create the
schema object that specifies the structure of the output BigQuery
table. Like the past section, we are still sampling the data set to
make sure our pipeline works before ramping up to the complete
data set. Once we’re confident in our pipeline, we can remove the
limit command and autoscale a cluster to complete the workload.

import argparse
from apache_beam.options.pipeline_options import SetupOptions

from apache_beam.io.gcp.bigquery import parse_table_schema_from_json
import json
query = """
,current_timestamp as time
,GENERATE_UUID() as guid
rand()
limit 100
"""
Next, we’ll define a DoFn class that implements the process function
and applies the sklearn model to individual records in the Natality
data set. One of the changes from before is that we now have an
init function, which we use to instantiate a set of fields. In order to
have references to the modules that we need to use in the process
function, we need to assign these as fields in the class, otherwise
the references will be undefined when running the function on dis-
tributed worker nodes. For example, we use self._pd to refer to
the Pandas module instead of pd. For the model, we’ll use lazy ini-
tialization to fetch the model from Cloud Storage once it’s needed.
While it’s possible to implement the setup function defined by the
DoFn interface to load the model, there are limitations on which
runners call this function.
class ApplyDoFn(beam.DoFn):
def __init__(self):
self._model = None

import pandas as pd
import pickle as pkl
self._storage = storage
self._pkl = pkl
self._pd = pd

if self._model is None:
bucket = self._storage.Client().get_bucket(
'dsp_model_store')
blob = bucket.get_blob('natality/sklearn-linear')
self._model =self._pkl.loads(blob.download_as_string())
new_x = self._pd.DataFrame.from_dict(element,
orient = "index").transpose().fillna(0)
weight = self._model.predict(new_x.iloc[:,1:8])[0]
return [ { 'guid': element['guid'], 'weight': weight,
'time': str(element['time']) } ]
Once the model object has been lazily loaded in the process func-
tion, it can be used to apply the linear regression model to the
input record. In Dataflow, records retrieved from BigQuery are re-
turned as a collection of dictionary objects and our process function
is responsible for operating on each of these dictionaries indepen-
dently. We first convert the dictionary to a Pandas dataframe and
then pass it to the model to get a predicted weight. The process
function returns a list of dictionary objects, which describe the
results to write to BigQuery. A list is returned instead of a dictio-
nary, because a process function in Beam can return zero, one, or
multiple objects.
An example element object passed to process function is shown in
the listing below. The object is a dictionary type, where the keys
are the column names of the query record and the values are the
record values.
{'year': 2001, 'plurality': 1, 'apgar_5min': 99, 'mother_age': 33,

'father_age': 40, 'gestation_weeks': 38, 'ever_born': 8,
'mother_married': 1, 'weight': 6.8122838958,
'time': '2019-12-14 23:51:42.560931 UTC',
'guid': 'b281c5e8-85b2-4cbd-a2d8-e501ca816363'}
To save the predictions to BigQuery, we need to define a schema

that defines the structure of the predictions table. We can do this
using a utility function that converts a JSON description of the ta-
ble schema into the schema object required by the Beam BigQuery
writer. To simplify the process, we can create a Python dictionary
object and use the dumps command to generate JSON.
schema = parse_table_schema_from_json(json.dumps({'fields':
[ { 'name': 'guid', 'type': 'STRING'},
{ 'name': 'weight', 'type': 'FLOAT64'},
{ 'name': 'time', 'type': 'STRING'} ]}))
The next step is to create the pipeline and define a DAG of Beam
operations. This time we are not providing input or output argu-
ments to the pipeline, and instead we are passing the input and
output destinations to the BigQuery operators. The pipeline has
three steps: read from BigQuery, apply the model, and write to
BigQuery. To read from BigQuery, we pass in the query and spec-
ify that we are using standard SQL. To apply the model, we use
our custom class for making predictions. To write the results, we
pass the schema and table name to the BigQuery writer, and spec-
ify that a new table should be created if necessary and that data
should be appended to the table if data already exists.
# set up pipeline options

FIGURE 7.4: The Natality predictions table on BigQuery.

data = p | 'Read from BigQuery' >> beam.io.Read(
beam.io.BigQuerySource(query=query, use_standard_sql=True))
scored = data | 'Apply Model' >> beam.ParDo(ApplyDoFn())
scored | 'Save to BigQuery' >> beam.io.Write(beam.io.BigQuerySink(
'weight_preds', 'dsp_demo', schema = schema,
create_disposition=beam.io.BigQueryDisposition.CREATE_IF_NEEDED,
write_disposition=beam.io.BigQueryDisposition.WRITE_APPEND))
The last step in the script is running the pipeline. While it is pos-
sible to run this complete code listing from Jupyter, the pipeline
will not be able to complete because the project parameter needs
to be passed as a command line argument to the pipeline.
# run the pipeline

result = p.run()
Before running the pipeline on Dataflow, it’s a best practice to

run the pipeline locally with a subset of data. In order to run
the pipeline locally, it’s necessary to specify the GCP project as a
command line argument, as shown below. The project parameter

is needed to read and write data with BigQuery. After running
the pipeline, you can validate that the workflow was successful
by navigating to the BigQuery UI and checking for data in the
destination table, as shown in Figure 7.4.
To run the pipeline on Cloud Dataflow, we need to pass a param-
eter that identifies the Dataflow Runner as the execution engine.
We also need to pass the project name and a staging location on
Cloud Storage. We now pass in a requirements file that identifies
the google-cloud-storage library as a dependency, and set a cluster
size limit using the max workers parameter. Once submitted, you
can view the progress of the job by navigating to the Dataflow UI
in the GCP console, as shown in Figure 7.5.
# running locally
python3 apply.py --project your_project_name
# running on GCP
echo $'google-cloud-storage==1.19.0' > reqs.txt
python3 apply.py \
--temp_location gs://dsp_model_store/tmp/ \
--requirements_file reqs.txt \
--maxNumWorkers 5
We can now remove the limit command from the query in the
pipeline and scale the workload to the full dataset. When running
the full-scale pipeline, it’s useful to keep an eye on the job to
make sure that the cluster size does not scale beyond expectations.
Setting the maximum worker count helps avoid issues, but if you
forget to set this parameter than the cluster size can quickly scale
and result in a costly pipeline run.
One of the potential issues with using Python for Dataflow
pipelines is that it can take awhile to initialize a cluster, because
each worker node will install the required libraries for the job from
FIGURE 7.5: Running the managed pipeline with autoscaling.
source, which can take a significant amount of time for libraries

such as Pandas. To avoid lengthy startup delays, it’s helpful to
avoid including libraries in the requirements file that are already
included in the Dataflow SDK1 . For example, Pandas 0.24.2 is in-
cluded with SDK version 2.16.0, which is a recent enough version
for this pipeline.
One of the useful aspects of Cloud Dataflow is that it is fully man-
aged, which means that it handles provisioning hardware, deals
with failures if any issues occur, and can autoscale to match de-
mand. Apache Beam is a great framework for data scientists, be-
cause it enables using the same tool for local testing and Google
Cloud deployments.
7.2.3 Datastore Publish

Publishing results to BigQuery is useful for ETLs and other appli-
cations that are part of batch data pipelines. However, it doesn’t
work well for use cases where applications need to retrieve a pre-
diction for users with low latency. GCP provides two NoSQL
databases that provide a solution for this use case, where you need
1
https://cloud.google.com/dataflow/docs/concepts/sdk-worker-dependencies
FIGURE 7.6: Publishing to BigQuery and Datastore.
to retrieve information for a specific user with minimal latency. In

this section we’ll explore Cloud Datastore, which provides some
querying capabilities for an application database.
We’ll build upon our prior pipeline and add an additional step
that publishes to Datastore while also publishing the predictions to
BigQuery. The resulting DAG is shown in Figure 7.6. The approach
we’ll use will write each entity to Datastore as part of the process
function. This is much slower than writing all of the entities as a
single Beam operation, but the Beam operator that performs this
step still requires Python 2.
To update the pipeline, we’ll define a new DoFn class and add this
as the last step in the pipeline, as shown in the code snippet be-
low. The init function loads the datastore module and makes it
referenceable as a field. The process function creates a key object
that is used to index the entity we want to store, which is similar
to a dictionary object. A Datastore entity is the base object used
to persist state in a Datastore database. We’ll use the guid as a
unique index for the entity and assign weight and time attributes
to the object. Once we’ve set the attributes, we use the put com-
mand to persist the object to Cloud Datastore. This last step can
take some time, which is why it is better to return a collection of
entities and perform the put step as part of a batch operation, if
supported.
FIGURE 7.7: The resulting entities in Cloud Datastore.
class PublishDoFn(beam.DoFn):
def __init__(self):
from google.cloud import datastore
self._ds = datastore

client = self._ds.Client()
key = client.key('natality-guid', element['guid'])
entity = self._ds.Entity(key)
entity['weight'] = element['weight']
entity['time'] = element['time']
client.put(entity)
scored | 'Create entities' >> beam.ParDo(PublishDoFn())
We can now rerun the pipeline to publish the predictions to both

BigQuery and Datastore. To validate that the model ran success-
fully, you can navigate to the Datastore UI in the GCP console
and inspect the entities, as shown in Figure 7.7.
7.3 Conclusion 199

client = datastore.Client()
query = client.query(kind='natality-guid')
query_iter = query.fetch()
for entity in query_iter:
print(entity)
break
It’s also possible to fetch the predictions published to Dataflow us-

ing Python. The code snippet above shows how to fetch all of the
model predictions and print out the first entity retrieved. A sam-
ple entity is shown below, which contains the guid as the unique
identifier and additional attributes for the weight and time of the
model prediction.
<Entity('natality-guid', '0046cdef-6a0f-4586-86ec-4b995cfc7c4e')
{'weight': 7.9434742419056,
'time': '2019-12-15 03:00:06.319496 UTC'}>
We now have a Dataflow pipeline that can scale to a large data

set and output predictions to analytical and application databases.
This means that other services can fetch these predictions to per-
sonalize products. Batch pipelines are one of the most common
ways that I’ve seen ML models productized in the gaming indus-
try, and Dataflow provides a great framework for enabling data
scientists to own more of the model production process.
7.3 Conclusion
Dataflow is a powerful data pipeline tool that enables data scien-
tists to rapidly prototype and deploy data processing workflows
that can apply machine learning algorithms. The framework pro-
vides a few basic operations that can be chained together to define
complex graphs of workflows. One of the key features of Dataflow

is that it builds upon an open source library called Apache Beam
that enables the workflows to be portable to other cloud environ-
ments.
In this chapter we built a batch model pipeline that fetched data
from BigQuery, applied a linear regression model, and then per-
sisted the predictions to BigQuery and Cloud Datastore. In the
next chapter we’ll explore a streaming version of this pipeline and
reuse portions of the current pipeline.
8
Streaming Model Workflows
Many organizations are now using streaming platforms in order

to build real-time data pipelines that transform streams of data
and move data between different components in a cloud envi-
ronment. These platforms are typically distributed and provide
fault-tolerance for streaming data. In addition to connecting dif-
ferent systems together, these tools also provide the ability to store
records and create event queues.
One of the most popular streaming platforms is Apache Kafka,
which is an open-source solution for providing message streaming
across public and private clouds. Kafka is a hosted solution that
requires provisioning and managing a cluster of machines in order
to scale. GCP provides a fully-managed streaming platform called
PubSub and AWS provides a managed solution called Kinesis. The
best option to use depends on your cloud platform, throughput and
latency requirements, and DevOps concerns.
With a streaming platform, you can pass data between different
components in a cloud environment, as well as external systems.
For example, many game companies are now using these platforms
to collect gameplay events from mobile games, where an event is
transmitted from the game client to a game server, and then passed
to the data platform as a stream. The message producer in this
case is the game server passing the message to the consumer, which
is the data platform that transforms and stores the event.
The connection to data science in production is that streaming
platforms can be used to apply ML models as a transform step in
a streaming pipeline. For example, you can set up a Python process
that reads in messages from a topic, applies a sklearn model, and
outputs the prediction to a new topic. This process can be part of
201
202 8 Streaming Model Workflows
a larger workflow that provides real-time ML predictions for users,

such as item recommendations in a mobile game. For the model
application step to scale to large volumes of messages, we’ll need to
use distributed systems such as Spark and Cloud Dataflow rather
than a single Python process.
While the model application step in a streaming model pipeline is
similar to setting up a Lambda or Cloud Function, which already
provides near real-time predictions, a key difference is the ease
of integrating with other components in the cloud platform. For
example, with Cloud Dataflow you can route the event to BigQuery
for storage as well as the model application step, which may push
the output to a new message consumer. Another benefit is that it
enables using distributed tools such as PySpark to handle requests
for model application, versus the endpoint based approaches that
service requests in isolation.
One of the benefits of using messaging systems in a cloud plat-
form is that it enables different tools and different programming
languages to communicate using standardized interfaces. We’ll fo-
cus on Python and PySpark in this book, but Java, Go, and many
other languages are supported by these platforms. In this chap-
ter, we’ll first use Apache Kafka to pass messages between differ-
ent Python processes and then consume, transform, and produce
new messages using PySpark Streaming. Next, we’ll use PubSub
on GCP to provide near real-time model predictions using Cloud
Dataflow in streaming mode.
8.1 Spark Streaming

Streaming data sets have been supported in Spark since version
0.7, but it was not until version 2.3 that a low-latency mode called
Structured Streaming was released. With structured streaming,
continuous processing can be used to achieve millisecond laten-
cies when scaling to high-volume workloads. The general flow with
structured streaming is to read data from an input stream, such
8.1 Spark Streaming 203
as Kafka, apply a transformation using Spark SQL, Dataframe

APIs, or UDFs, and write the results to an output stream. Spark
Streaming also works with managed streaming platforms including
PubSub and Kinesis, and other frameworks in the Apache ecosys-
tem including Flume.
In this section, we’ll first set up a Kafka instance and then produce
and consume messages on the same machine using Kafka. Next,
we’ll show how to consume messages from Kafka using the read-
Stream function in PySpark, and then build a streaming pipeline
that applies a sklearn model.
8.1.1 Apache Kafka

Kafka is an open-source streaming platform that was incubated
at LinkedIn. It is designed to handle real-time data streams that
are high throughput and low latency. It is written in Java and
Scala, but supports a range of programming languages for pro-
ducing and consuming streams through standardized APIs. The
platform can scale to large data sets by using horizontal scaling
and partitioning to distribute workloads across a cluster of servers
called brokers. While open-source Kafka is a hosted solution for
message streaming, some cloud providers now offer fully-managed
versions of Kafka, such as Amazon’s MSK offering.
To show how Kafka can be integrated into a streaming workflow,
we’ll use a single-node setup to get up and running. For a produc-
tion environment, you’ll want to set up a multi-node cluster for
redundancy and improved latency. Since the focus of this chap-
ter is model application, we won’t dig into the details of setting
up Kafka for high-availability, and instead recommend managed
solutions for small teams getting started. To install Kafka, it’s use-
ful to browse to the website1 and find the most recent release. In
order to install Kafka, we first need to install Java, and then down-
load and extract the Kafka release. The steps needed to set up a
single-node Kafka instance on an EC2 machine are shown in the
1
https://kafka.apache.org/quickstart
snippet below. We’ll also install a library for working with Kafka
in Python called kafka-python.
sudo yum install -y java

pip install --user kafka-python
wget http://mirror.reverse.net/pub/apache/kafka/2.4.0/
kafka_2.12-2.4.0.tgz
tar -xzf kafka_2.12-2.4.0.tgz
cd kafka_2.12-2.4.0
bin/zookeeper-server-start.sh config/zookeeper.properties
# new terminal
bin/kafka-server-start.sh config/server.properties
# new terminal
bin/kafka-topics.sh --create --bootstrap-server localhost:9092
--replication-factor 1 --partitions 1 --topic dsp
# output
[2019-12-18 10:50:25] INFO Log partition=dsp-0, dir=/tmp/kafka-logs
Completed load of log with 1 segments, log start offset 0 and
log end offset 0 in 56 ms (kafka.log.Log)
When setting up Kafka, we’ll need to spawn three separate pro-

cesses to run dependencies, start the Kafka service, and create a
new topic for publishing messages. The snippet above runs the
following processes:
• Zookeeper: An Apache project that provides configuration and
service discovery for distributed systems.
• Kafka Launches the bootstrap service that enables setting up
Kafka topics and using streaming APIs.
• Topics: Creates a new topic called “dsp”.
The Zookeeper and Kafka tasks are long-running processes that will
continue to execute until terminated, while the Topics process will
shutdown once the new Kafka topic is set up. The output at the
bottom of the snippet shows the output from running this com-
mand, which will be displayed in terminal running the Kafka pro-

cess. In this configuration, we are setting up a single partition for
the topic with no replication. We now have a single-node Kafka
cluster set up for testing message streaming.
The first API that we’ll explore is the Producer API, which enables
processes to publish a message to a topic. To publish a message to
our Kafka server, we create a producer object by passing in an IP
address and a serialization function, which specifies how to encode
Python objects into strings that can be passed to the Kafka server.
The Python snippet below shows how to create the producer and
send a dictionary object as a message to the server, publishing
the message to the dsp topic. The dict object contains hello and
time keys. If we run this code, the message should be successfully
transmitted to the server, but there will not yet be a consumer to
process the message.
from kafka import KafkaProducer

from json import dumps
import time
producer = KafkaProducer(bootstrap_servers=['localhost:9092'],
value_serializer=lambda x: dumps(x).encode('utf-8'))
data = {'hello' : 'world', 'time': time.time()}

producer.send('dsp', data)
To set up a process for consuming the message, we’ll explore the

Consumer API, which is used to read in streams of data. The Python
snippet below shows how to create a consumer object that connects
to the Kafka server and subscribes to the dsp topic. The consumer
object returned is iterable and can be used in combination with a
for loop in order to process messages. In the example below, the
for loop will suspend execution until the next message arrives and
continue iterating until the process is terminated. The value ob-
ject will be a Python dictionary that we passed from the producer,
while the deserializer function defines how to transform strings to
Python objects. This approach works fine for small-scale streams,

but with a larger data volume we also want to distribute the mes-
sage processing logic, which we’ll demonstrate with PySpark in the
next section.
from kafka import KafkaConsumer

from json import loads
consumer = KafkaConsumer('dsp',
bootstrap_servers=['localhost:9092'],
value_deserializer=lambda x: loads(x.decode('utf-8')))
for x in consumer:
print(x.value)
Now that we have Python scripts for producing and consuming

messages, we can test message streaming with Kafka. First, run
the Consumer script in a Jupyter notebook, and then run the Pro-
ducer script in a separate notebook. After running the producer
cell multiple times, you should see output from the consumer cell
similar to the results shown below.
{'hello': 'world', 'time': 1576696313.876075}

{'hello': 'world', 'time': 1576696317.435035}
{'hello': 'world', 'time': 1576696318.219239}
We can now use Kafka to reliably pass messages between differ-

ent components in a cloud deployment. While this section used a
test configuration for spinning up a Kafka service, the APIs we
explored apply to production environments with much larger data
volumes. In the next section, we’ll explore the Streams API which
is used to process streaming data, such as applying an ML model.
8.1.2 Sklearn Streaming

To build an end-to-end streaming pipeline with Kafka, we’ll lever-
age Spark streaming to process and transform data as it arrives.
The structured streaming enhancements introduced in Spark 2.3

enable working with dataframes and Spark SQL while abstracting
away many of the complexities of dealing with batching and pro-
cessing data sets. In this section we’ll set up a PySpark streaming
pipeline that fetches data from a Kafka topic, applies a sklearn
model, and writes the output to a new topic. The entire workflow
is a single DAG that continuously runs and processes messages
from a Kafka service.
In order to get Kafka to work with Databricks, we’ll need to edit
the Kafka configuration to work with external connections, since
Databrics runs on a separate VPC and potentially separate cloud
than the Kafka service. Also, we previously used the bootstrap
approach to refer to brokers using localhost as the IP. On AWS,
the Kafka startup script will use the internal IP to listen for con-
nection, and in order to enable connections from remote machines
we’ll need to update the configuration to use the external IP, as
shown below.
vi config/server.properties
advertised.listeners=PLAINTEXT://{external_ip}:9092
After making this configuration change, you’ll need to restart

the Kafka process in order to receive inbound connections from
Databricks. You’ll also need to enable inbound connections from
remote machines, by modifying the security group, which is cov-
ered in Section 1.4.1. Port 9092 needs to be open for the Spark
nodes that will be making connections to the Kafka service.
We’ll also set up a second topic, which is used to publish the results
of the model application step. The PySpark workflow we will set
up will consume messages from a topic, apply a sklearn model, and
then write the results to a separate topic, called preds. One of the
key benefits of this workflow is that you can swap out the pipeline
that makes predictions without impacting other components in
the system. This is similar to components in a cloud workflow
calling out to an endpoint for predictions, but instead of changing
the configuration of components calling endpoints to point to new
endpoints, we can seamlessly swap in new backend logic without

impacting other components in the workflow.
bin/kafka-topics.sh --create --bootstrap-server localhost:9092

--replication-factor 1 --partitions 1 --topic preds
It’s a good practice to start with a basic workflow that simply

consumes messages before worrying about how to build out a pre-
dictive modeling pipeline, especially when working with streaming
data. To make sure that we’ve correctly set up Kafka for remote
connections with Databricks, we can author a minimal script that
consumes messages from the stream and outputs the results, as
shown in the PySpark snippet below. Databricks will refresh the
output on a regular interval and show new data in the output table
as it arrives. Setting the startingOffsets value to earliest means
that we’ll backload data from the last Kafka checkpoint. Remov-
ing this setting will mean that only new messages are displayed in
the table.
df = spark .readStream.format("kafka")
.option("kafka.bootstrap.servers", "{external_ip}:9092")
.option("subscribe", "dsp")
.option("startingOffsets", "earliest").load()
display(df)
Getting Databricks to communicate with the Kafka service can

be one of the main challenges in getting this sample pipeline to
work, which is why I recommend starting with a minimal PySpark
script. It’s also useful to author simple UDFs that process the
value field of the received messages to ensure that the decoded
message in PySpark matches the encoded data from the Python
process. Once we can consume messages, we’ll use a UDF to apply
a sklearn model, where UDF refers to a Python function and not
a Pandas UDF. As a general practice, it’s good to add checkpoints
to a Spark workflow, and the snippet above is a good example for
checking if the data received matches the data transmitted.
For the Spark streaming example, we’ll again use the Games data
set, which has ten attributes and a label column. In this workflow,
we’ll send the feature vector to the streaming pipeline as input,
and output an additional prediction column as the output. We’ll
also append a unique identifier, as shown in the Python snippet
below, in order to track the model applications in the pipeline.
The snippet below shows how to create a Python dict with the ten
attributes needed for the model, append a GUID to the dictionary,
and send the object to the streaming model topic.
from kafka import KafkaProducer

from json import dumps
import time
import uuid
producer = KafkaProducer(bootstrap_servers= ['{external_ip}:9092'],

value_serializer=lambda x: dumps(x).encode('utf-8'))
data = { 'G1': 1, 'G2': 0, 'G3': 0, 'G4': 0, 'G5': 0,

'G6': 0, 'G7': 0, 'G8': 0, 'G9': 0, 'G10': 0,
'User_ID': str(uuid.uuid1())}
result = producer.send('dsp', data)
result.get()
To implement the streaming model pipeline, we’ll use PySpark

with a Python UDF to apply model predictions as new elements
arrive. A Python UDF operates on a single row, while a Pandas
UDF operates on a partition of rows. The code for this pipeline
is shown in the PySpark snippet below, which first trains a model
on the driver node, sets up a data sink for a Kafka stream, defines
a UDF for applying an ML model, and then publishes the scores
to a new topic as a pipeline output.
from pyspark.sql.types import StringType

import json
import pandas as pd
# build a logistic regression model

model.fit(gamesDF.iloc[:,0:10], gamesDF['label'])
# define the UDF for scoring users

def score(row):
d = json.loads(row)
p = pd.DataFrame.from_dict(d, orient = "index").transpose()
pred = model.predict_proba(p.iloc[:,0:10])[0][0]
result = {'User_ID': d['User_ID'], 'pred': pred }
return str(json.dumps(result))
# read from Kafka

df = spark.readStream.format("kafka")
.option("subscribe", "dsp").load()
# select the value field and apply the UDF

df = df.selectExpr("CAST(value AS STRING)")
score_udf = udf(score, StringType())
df = df.select( score_udf("value").alias("value"))
# Write results to Kafka

query = df.writeStream.format("kafka")
.option("topic", "preds")
.option("checkpointLocation", "/temp").start()
The script first trains a logistic regression model using data fetched
from GitHub. The model object is created on the driver node, but
is copied to the worker nodes when used by the UDF. The next
step is to define a UDF that we’ll apply to streaming records in the
pipeline. The Python UDF takes a string as input, converts the

string to a dictionary using the json library, and then converts the
dictionary into a Pandas dataframe. The dataframe is passed to
the model object and the UDF returns a string representation of a
dictionary object with User_ID and pred keys, where the prediction
value is the propensity of the user to purchase a specific game.
The next three steps in the pipeline define the PySpark streaming
workflow. The readStream call sets up the connection to the Kafka
broker and subscribes to the dsp topic. Next, a select statement
is used to cast the value column of streaming records to a string
before passing the value to the UDF, and then creating a new
dataframe using the result of the Python UDF. The last step writes
the output dataframe to the preds topic, using a local directory as
a checkpoint location for Kafka. These three steps run as part of a
continuous processing workflow, where the steps do not complete,
but instead suspend execution until new data arrives. The result
is a streaming DAG of operations that processes data as it arrives.
When running a streaming pipeline, Databricks will show details
about the workflow below the cell, as shown in Figure 8.1. The
green icon identifies that this is a streaming operation that will
continue to execute until terminated. There are also charts that
visualize data throughput and latency. For a production pipeline,
it’s useful to run code using orchestration tools such as Airflow
with the Databricks operator, but the notebook environment does
provide a useful way to run and debug streaming pipelines.
Now that we are streaming model predictions to a new topic, we’ll
need to create a new consumer for these messages. The Python
snippet below shows to consume messages from the broker for the
new predictions topic. The only change from the prior consumer
is the IP address and the deserializer function, which no longer
applies an encoding before converting the string to a dictionary.
from kafka import KafkaConsumer

from json import loads
FIGURE 8.1: Visualizing stream processing in Databricks.
consumer = KafkaConsumer('preds',
bootstrap_servers=['{external_ip}:9092'],
value_deserializer=lambda x: loads(x))
for x in consumer:
print(x.value)
We now have everything in place in order to test out the streaming

pipeline with Spark streaming and Kafka. First, run the PySpark
pipeline in a Databricks cell. Next, run the consumer script in a
Jupyter notebook. To complete the workflow, run the producer
script in a separate Jupyter notebook to pass a message to the
pipeline. The result should be a prediction dictionary printed to
the console of the consumer notebook, as shown below.
{'User_ID': '4be94cd4-21e7-11ea-ae04-8c8590b3eee6',
'pred': 0.9325488640736544}
We now have a PySpark streaming pipeline that applies model

predictions with near real-time performance. There’s additional
tuning that we can perform to get this latency to within one mil-
lisecond, which is useful for a variety of model and web applica-
tions. The benefit of using Spark to perform model application is
that we can scale the cluster to match demand and can swap in
8.2 Dataflow Streaming 213
new pipelines as needed to provide model updates. Spark stream-

ing was initially a bit tricky to get up and running, but the recent
enhancements have made it much easier to get working with model
application pipelines. In this pipeline we used a simple regression
model, but streaming workflows can also be used for deep learning
tasks, such as image classification2 .
8.2 Dataflow Streaming

In Chapter 7 we explored using Cloud Dataflow to create a batch
model pipeline, and authored a DoFn to apply a sklearn model.
Dataflow can also be used to build streaming model pipelines, by
setting a configuration flag. One of the features of Dataflow is
that many components can be reused across batch and streaming
workflows, and in this section we’ll use the same model application
class that we defined in the last chapter.
When working with Dataflow in streaming mode, you can use a
combination of streaming and batch data sources and streaming
data sinks when defining a DAG. For example, you can use the
approach from the prior section to read from a Kafka stream, apply
a model with a DoFn function, and write the results to another
stream as a data sink. Some of the GCP systems that work with
Dataflow, such as BigQuery, can be used as both a batch and
streaming data sink.
In this section, we’ll build a streaming pipeline with Dataflow that
streams in messages from PubSub, applies a sklearn model, and
publishes the results to Cloud Datastore. This type of pipeline is
useful for updating a user profile based on real-time data, such as
predicting if a user is likely to make a purchase.
2
https://www.youtube.com/watch?v=xwQwKW-cerE
8.2.1 PubSub
PubSub is a fully-managed streaming platform available on GCP.
It provides similar functionality to Kafka for achieving high
throughput and low latency when handling large volumes of mes-
sages, but reduces the amount of DevOps work needed to maintain
the pipeline. One of the benefits of PubSub is that the APIs map
well to common use cases for streaming Dataflow pipelines.
One of the differences from Kafka is that PubSub uses separate
concepts for producer and consumer data sources. In Kafka, you
can publish and subscribe to a topic directly, while in PubSub con-
sumers subscribe to subscriptions rather than directly subscribing
to topics. With PubSub, you first set up a topic and then create
one or more subscriptions that listen on this topic. To create a
topic with PubSub, perform the following steps:
1. Browse to the PubSub UI in the GCP Console

2. Click “Create Topic”
3. Enter “natality” for the topic ID, and click “Create Topic”
The result of performing these actions is that we now have a topic

called natality that we can use for publishing messages. Next, we’ll
create a subscription that listens for messages on this topic by
performing these steps:
1. Click on Subscriptions in the navigation pane

2. Click “Create Subscription”
3. Assign a subscription ID “dsp”
4. Select the “natality” topic
We now have a topic and subscription set up for streaming mes-

sages in a pipeline. Before setting up a Dataflow pipeline, we’ll first
create message consumers and producers in Python. The code snip-
pet below shows how to read messages from the subscription using
the Google Cloud library. We first create a subscriber client, then
set up the subscription and assign a callback function. Unlike the
Kafka approach, which returns an iterable object, PubSub uses
a callback pattern where you provide a function that is used to

process messages as they arrive. In this example we simply print
the data field in the message and then acknowledge that the mes-
sage has been received. The for loop at the bottom of the code
block is used to keep the script running, because none of the other
commands suspend when executed.
import time
from google.cloud import pubsub_v1
subscriber = pubsub_v1.SubscriberClient()
subscription_path = subscriber.subscription_path(
"your_project_name", "dsp")
def callback(message):
print(message.data)
message.ack()
subscriber.subscribe(subscription_path, callback=callback)
while True:
time.sleep(10)
We’ll use the same library to create a message producer in Python.

The code snippet below shows how to use the Google Cloud library
to create a publishing client, set up a connection to the topic, and
publish a message to the dsp topic. For the producer, we need to
encode the message in utf-8 format before publishing the message.
publisher = pubsub_v1.PublisherClient()
topic_path = publisher.topic_path("your_project_name8", "natality")
data = "Hello World!".encode('utf-8')

publisher.publish(topic_path, data=data)
To test out the pipeline, first run the consumer in a Jupyter note-
book and then run the producer in a separate Jupyter notebook.
The result should be that the consumer cell outputs "Hello World"
to the console after receiving a message. Now that we have tested
out basic functionality with PubSub, we can now integrate this
messaging platform into a streaming Dataflow pipeline.
8.2.2 Natality Streaming

PubSub can be used to provide data sources and data sinks within
a Dataflow pipeline, where a consumer is a data source and a
publisher is a data sink. We’ll reuse the Natality data set to create
a pipeline with Dataflow, but for the streaming version we’ll use a
PubSub consumer as the input data source rather than a BigQuery
result set. For the output, we’ll publish predictions to Datastore
and reuse the publish DoFn from the previous chapter.

import argparse
from apache_beam.io.gcp.bigquery
import parse_table_schema_from_json
import json
class ApplyDoFn(beam.DoFn):
def __init__(self):
self._model = None
import pandas as pd
import pickle as pkl
import json as js
self._storage = storage
self._pkl = pkl
self._pd = pd
self._json = js

if self._model is None:
bucket = self._storage.Client().get_bucket(
'dsp_model_store')
blob = bucket.get_blob('natality/sklearn-linear')
self._model =self._pkl.loads(blob.download_as_string())
element = self._json.loads(element.decode('utf-8'))
new_x = self._pd.DataFrame.from_dict(element,
orient = "index").transpose().fillna(0)
weight = self._model.predict(new_x.iloc[:,1:8])[0]
return [ { 'guid': element['guid'], 'weight': weight,
'time': str(element['time']) } ]
The code snippet above shows the function we’ll use to perform
model application in the streaming pipeline. This function is the
same as the function we defined in Chapter 7 with one modification,
the json.loads function is used to convert the passed in string into
a dictionary object. In the previous pipeline, the elements passed
in from the BigQuery result set were already dictionary objects,
while the elements passed in from the PubSub consumer are string
objects. We’ll also reuse the DoFn function the past chapter which
publishes elements to Datastore, listed in the snippet below.
class PublishDoFn(beam.DoFn):
def __init__(self):
self._ds = datastore

client = self._ds.Client()
key = client.key('natality-guid', element['guid'])
entity = self._ds.Entity(key)
entity['weight'] = element['weight']
entity['time'] = element['time']
client.put(entity)
Now that we have defined the functions for model application

and publishing to Datastore, we can build a streaming DAG with
dataflow. The Python snippet below shows how to build a Dataflow
pipeline that reads in a message stream from the natality subscrip-
tion, applies the model application function, and then publishes
the output to the application database.
# set up pipeline parameters

# define the topics

topic = "projects/{project}/topics/{topic}"
topic = topic.format(project="your_project_name", topic="natality")

lines = p | 'Read PubSub' >> beam.io.ReadFromPubSub(topic=topic)
scored = lines | 'apply' >> beam.ParDo(ApplyDoFn())
scored | 'Create entities' >> beam.ParDo(PublishDoFn())
# run the pipeline

result = p.run()
The code does not explicitly state that this is a streaming pipeline,
and the code above can be executed in a batch or streaming mode.
In order to run this pipeline as a streaming Dataflow deployment,
we need to specify the streaming flag as shown below. We can first
test the pipeline locally before deploying the pipeline to GCP. For
a streaming pipeline, it’s best to use GCP deployments, because
the fully-managed pipeline can scale to match demand, and the
FIGURE 8.2: The Dataflow DAG with streaming metrics.
platform will handle provisioning hardware and provide fault tol-

erance. A visualization of the Dataflow pipeline running on GCP
is shown in Figure 8.2.
python3 natality.py --streaming
To test out the pipeline, we’ll need to pass data to the dsp topic
which is forwarded to the natality subscription. The code snippet
below shows how to pass a dictionary object to the topic using
Python and the Google Cloud library. The data passed to Pub-
Sub represents a single record in the BigQuery result set from the
previous chapter.
import json
import time
data = json.dumps({'year': 2001, 'plurality': 1,

'apgar_5min': 99, 'mother_age': 33,
'father_age': 40, 'gestation_weeks': 38, 'ever_born': 8,
'mother_married': 1, 'weight': 6.8122838958,
'time': str(time.time()),
'guid': 'b281c5e8-85b2-4cbd-a2d8-e501ca816363'}
).encode('utf-8')
publisher = pubsub_v1.PublisherClient()
FIGURE 8.3: The prediction output pushed to Cloud Datastore.
topic_path = publisher.topic_path("your_project_name", "natality")

publisher.publish(topic_path, data=data)
The result of passing data to the topic should be an updated entry

in Datastore, which provides a model prediction for the passed in
GUID. As more data is passed to the pipeline, additional entries
will be added to the data set. A sample output of this pipeline is
shown in Figure 8.3, which displays the predicted weight for one
of the records passed to the pipeline. To run the pipeline on GCP,
run the following statement on the command line.
python3 natality.py --streaming

--temp_location gs://dsp_model_store/tmp/ \
We now have a Dataflow streaming pipeline running in a fully-

managed environment. We can use PubSub to interface the
pipeline with other components in a cloud deployment, such as
a data platform that receives real-time data from mobile appli-
cations. With Dataflow, many components can be reused across
batch and streaming model pipelines, which makes it a flexible
tool for building production pipelines. One factor to consider when
using Dataflow for streaming pipelines is that costs can be much
8.4 Conclusion 221
larger when using streaming versus batch operations, such as writ-

ing to BigQuery3 .
8.3 Conclusion
Streaming model pipelines are useful for systems that need to ap-
ply ML models in real-time. To build these types of pipelines, we
explored two message brokers that can scale to large volumes of
events and provide data sources and data sinks for these pipelines.
Streaming pipelines often constrain the types of operations you
can perform, due to latency requirements. For example, it would
be challenging to build a streaming pipeline that performs feature
generation on user data, because historic data would need to be re-
trieved and combined with the streaming data while maintaining
low latency. There are patterns for achieving this type of result,
such as precomputing aggregates for a user and storing the data
in an application database, but it can be significantly more work
getting this type of pipeline to work in a streaming mode versus a
batch mode.
We first explored Kafka as a streaming message platform and built
a real-time pipeline using structure streaming and PySpark. Next,
we built a steaming Dataflow pipeline reusing components from
the past chapter that now interface with the PubSub streaming
service. Kafka is typically going to provide the best performance
in terms of latency between these two message brokers, but it takes
significantly more resources to maintain this type of infrastructure
versus using a managed solution. For small teams getting started,
PubSub or Kinesis provide great options for scaling to match de-
mand while reducing DevOps support.
3
https://labs.spotify.com/2017/10/16/
8.4 Thank You

Writing this book has been a great experience for getting hands-on
with many of the tools that I advocate for data scientists to learn
and add to their toolbox. Thank you for reading through this book
to the end.
Given the breadth of tools covered in this book, I wasn’t able to
provide much depth on any particular topic. The next step for read-
ers is to choose a topic introduced in this text and find resources
to learn about the topic in more depth than can be covered here.
Data science is a rapidly evolving field, and that means that the
contents of this book will become outdated as cloud platforms
evolve and libraries are updated. While I was authoring this text,
substantial updates were released for TensorFlow which impacted
the later chapters. The takeaway is that keeping up with data
science as a discipline requires ongoing learning and practice.
Bibliography
Chollet, F. (2017). Deep Learning with Python. Manning, 1st

edition. ISBN 978-1617294433.
Géron, A. (2017). Hands-On Machine Learning with Scikit-Learn
and TensorFlow. O’Reilly Media, 1st edition. ISBN 978-
1491962299.
Grus, J. (2015). Data Science from Scratch. O’Reilly Media, 1st
edition. ISBN 978-1491901427.
Karau, H., Konwinski, A., Wendell, P., and Zaharia, M. (2015).
Learning Spark: Lightning-Fast Big Data Analysis. O’Reilly Me-
dia, 1st edition. ISBN 978-1449358624.
McKinney, W. (2017). Python for Data Analysis. O’Reilly Media,
2nd edition. ISBN 978-1491957660.
Weber, B. (2018). Data Science for Startups. Kindle Direct, 1st
edition. ISBN 978-1983057977.
Xie, Y. (2015). Dynamic Documents with R and knitr. Chapman
and Hall/CRC, Boca Raton, Florida, 2nd edition. ISBN 978-
1498716963.
Xie, Y. (2019). bookdown: Authoring Books and Technical Docu-
ments with R Markdown. R package version 0.16.
223

Ben G Weber - Data Science in Production - Building Scalable Model Pipelines With Python-Independently Published (2020)

Uploaded by

Copyright:

Available Formats

Ben G Weber - Data Science in Production - Building Scalable Model Pipelines With Python-Independently Published (2020)

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ben G Weber - Data Science in Production - Building Scalable Model Pipelines With Python-Independently Published (2020)

Uploaded by

Copyright:

Available Formats

Ben G.

Data Science in Production:

2 Models as Web Endpoints 33

2.3 Model Endpoints . . . . . . . . . . . . . . . . . . 43

3 Models as Serverless Functions 57

4 Containers for Reproducible Models 87

5 Workflow Tools for Model Pipelines 109

5.3 Workflow Tools . . . . . . . . . . . . . . . . . . . 120

6 PySpark for Batch Pipelines 129

7 Cloud Dataflow for Batch Modeling 181

7.2.2 BigQuery Publish . . . . . . . . . . . . . . 190

8 Streaming Model Workflows 201

0.2 Book Contents

1. Introduction: This chapter will motivate the use of

3. Models as Serverless Functions: This chapter will

After working through this material, readers should have hands-on

0.3 Code Examples

Due to formatting restrictions, many of the code snippets in this

Putting predictive models into production is one of the most di-

models, lifetime-value predictions, and recommendation systems

1.1 Applied Data Science

such as deep learning, and scalable computing environments, such

1.2 Python for Scalable Compute

order to directly collaborate with engineering teams on projects. I

• Learn the ecosystem, not the language: I didn’t limit myself

1.3 Cloud Environments

1.3.1 Amazon Web Services (AWS)

across a cluster of machines, and vendors such as Databricks pro-

1.3.2 Google Cloud Platform (GCP)

1.4 Coding Environments

1.4.1 Jupyter on EC2

FIGURE 1.1: Public and Private IPs on EC2.

1. Under “Find Services”, search for EC2

FIGURE 1.2: SSH connection settings.

as shown in Figure 1.1. To connect to the machine via SSH we’ll

sudo yum install -y python37

The two version commands are to confirm that the machine is

1. Select your EC2 instance

We can now run and connect to Jupyter on the EC2 machine. To

jupyter notebook --ip 172.31.53.82

FIGURE 1.3: Jupyter Notebook on EC2.

# Swap Private IP with Public IP

pip install --user pandas

Once scikit-learn is installed, we can switch back to the Jupyter

from sklearn.datasets import load_boston

FIGURE 1.4: Boston Housing data set.

data, target = load_boston(True)

The second approach we’ll use to load a data set is to fetch it

Both of these approaches are similar to downloading CSV files and

1.5.1 BigQuery to Pandas

FIGURE 1.5: Game Purchases data set.

a GCP project, installing the prerequisite Python libraries, setting

pip install --user google-cloud-bigquery

gcloud config set project project_name

You’ll need to substitute project_name with your project name,

from google.cloud import bigquery

FIGURE 1.6: Previewing the BigQuery data set.

1.5.2 Kaggle to Pandas