Cloud Native Applications

Selected by Michael Wittig

and Andreas Wittig

Manning Author Picks

Copyright 2016 Manning Publications

To pre-order or learn more about these books go to www.manning.com

Licensed to Marocco Aurélien <[email protected]>

For online information and ordering of these and other Manning books, please visit
www.manning.com. The publisher offers discounts on these books when ordered in quantity.

For more information, please contact

Special Sales Department

Manning Publications Co.
20 Baldwin Road
PO Box 761
Shelter Island, NY 11964
Email: [email protected]

©2016 by Manning Publications Co. All rights reserved.

No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in

any form or by means electronic, mechanical, photocopying, or otherwise, without prior written
permission of the publisher.

Many of the designations used by manufacturers and sellers to distinguish their products are
claimed as trademarks. Where those designations appear in the book, and Manning
Publications was aware of a trademark claim, the designations have been printed in initial caps
or all caps.

Recognizing the importance of preserving what has been written, it is Manning’s policy to have
the books we publish printed on acid-free paper, and we exert our best efforts to that end.
Recognizing also our responsibility to conserve the resources of our planet, Manning books
are printed on paper that is at least 15 percent recycled and processed without the use of
elemental chlorine.

Manning Publications Co.

20 Baldwin Road Technical
PO Box 761
Shelter Island, NY 11964

Cover designer: Leslie Haimes

ISBN 9781617294310
Printed in the United States of America
1 2 3 4 5 6 7 8 9 10 - EBM - 21 20 19 18 17 16

Licensed to Marocco Aurélien <[email protected]>

 contents
introduction iv

AMAZON WEB SERVICES IN ACTION 1

What is Amazon Web Services?
Chapter 1 from Amazon Web Services in Action 2

DOCKER IN PRACTICE 34
Discovering Docker
Chapter 1 from Docker in Practice 35

MESOS IN ACTION 52
Introducing Mesos
Chapter 1 from Mesos in Action 53

RABBITMQ IN DEPTH 68
Foundational RabbitMQ
Chapter 1 from RabbitMQ in Depth 69

NETTY IN ACTION 84
Case studies, part 1
Chapter 14 from Netty in Action 85
index 109

iii

Licensed to Marocco Aurélien <[email protected]>

 introduction
As engineers, we always aim to improve the feedback loop between us and our custom-
ers. This is one reason why we started to use Amazon Web Services four years ago.
Being able to automate every step from making changes in our source code to deploy-
ing the infrastructure needed to operate our applications was mind blowing. As cloud
consultants, we observe that being able to automate and constantly adjust infrastruc-
ture is one important reason to use cloud computing for our clients. Another reason
why we are focusing on cloud native applications entirely: developing and operating
systems that are able to recover from failure automatically was never easier. High Avail-
ability is becoming the new normal.
So what has changed? Cloud providers are offering the needed infrastructure for a
very reasonable price. AWS, for example, is providing multiple data centers per region
that you can use in parallel. Technologies like messaging systems and load balancers
allow you to decouple different parts of your system and plan for failure.
We encourage you to read these chapters we’ve selected from several Manning
books and learn how to gain value from cloud computing for your company. This
book will guide you through five topics giving you insights into the world of cloud
computing. You’ll learn how to use Amazon Web Services, one of the most important
public cloud providers. You’ll get to know to Docker and Mesos as well. Both tools
help you to automate and manage your cloud infrastructure. Developing software will
generate the biggest value if you adopt your architecture and processes. Decoupling
different parts of your system is possible by using a messaging infrastructure, which
you’ll learn about in the RabbitMQ chapter. In addition, you’ll discover different use
cases for Netty, a framework that helps you to build solid networking communication
into your applications for the cloud.

iv

Licensed to Marocco Aurélien <[email protected]>

 Amazon Web Services
in Action

W hat is the biggest advantage of using Amazon Web Services (AWS)? For
us it’s being able to automate every part of your cloud infrastructure. AWS offers
an API and lots of tools to launch, configure, modify, and delete computing,
storage and networking infrastructure. Our book, Amazon Web Services in Action,
provides a deep introduction into the most important services and architecture
principles. Chapter 1 answers the question: What is Amazon Web Services? You’ll
learn about the concepts behind AWS and gain a brief overview of what you can
do with AWS.

Licensed to Marocco Aurélien <[email protected]>

 Chapter 1 from Amazon Web Services in Action by
Michael Wittig and Andreas Wittig.

What is
Amazon Web Services?

This chapter covers

Overview of Amazon Web Services
Benefits of using Amazon Web Services
Examples of what you can do with Amazon Web
Services
Creating and setting up an Amazon Web Services
account

Amazon Web Services (AWS) is a platform of web services offering solutions for
computing, storing, and networking, at different layers of abstraction. You can use
these services to host web sites, run enterprise applications, and mine tremendous
amounts of data. The term web service means services can be controlled via a web
interface. The web interface can be used by machines or by humans via a graphical
user interface. The most prominent services are EC2, which offers virtual servers,
and S3, which offers storage capacity. Services on AWS work well together; you can

Licensed to Marocco Aurélien <[email protected]>

 What is cloud computing? 3

use them to replicate your existing on-premises setup or design a new setup from
scratch. Services are charged for on a pay-per-use pricing model.
As an AWS customer, you can choose among different data centers. AWS data cen-
ters are distributed in the United States, Europe, Asia, and South America. For exam-
ple, you can start a virtual server in Japan in the same way you can start a virtual server
in Ireland. This enables you to serve customers worldwide with a global infrastructure.
The map in figure 1.1 shows the data centers available to all customers.

Which hardware powers AWS?

AWS keeps secret the hardware used in its data centers. The scale at which AWS
operates computing, networking, and storage hardware is tremendous. It probably
uses commodity components to save money compared to hardware that charges
extra for a brand name. Handling of hardware failure is built into real-world processes
and software.1
AWS also uses hardware especially developed for its use cases. A good example is
the Xeon E5-2666 v3 CPU from Intel. This CPU is optimized to power virtual servers
from the c4 family.

In more general terms, AWS is known as a cloud computing platform.1

1.1 What is cloud computing?

Almost every IT solution is labeled with the term cloud computing or just cloud nowa-
days. A buzzword may help to sell, but it’s hard to work with in a book.
Cloud computing, or the cloud, is a metaphor for supply and consumption of IT
resources. The IT resources in the cloud aren’t directly visible to the user; there are
layers of abstraction in between. The level of abstraction offered by the cloud may vary

Germany
Ireland
U.S. West 1

U.S. West 2 U.S. East Japan

Singapore

Australia
Brazil

Figure 1.1 AWS data center locations

1
Bernard Golden, “Amazon Web Services (AWS) Hardware,” For Dummies, http://mng.bz/k6lT.

Licensed to Marocco Aurélien <[email protected]>

4 CHAPTER 1 What is Amazon Web Services?

from virtual hardware to complex distributed systems. Resources are available on

demand in enormous quantities and paid for per use.
Here's a more official definition from the National Institute of Standards and
Technology:
Cloud computing is a model for enabling ubiquitous, convenient, on-demand network
access to a shared pool of configurable computing resources (e.g., networks, servers,
storage, applications, and services) that can be rapidly provisioned and released with
minimal management effort or service provider interaction.
—The NIST Definition of Cloud Computing,
National Institute of Standards and Technology

Clouds are often divided into the following types:

Public —A cloud managed by an organization and open to use by the general
public
Private —A cloud that virtualizes and shares the IT infrastructure within a single
organization
Hybrid —A mixture of a public and a private cloud
AWS is a public cloud. Cloud computing services also have several classifications:
Infrastructure as a service (IaaS) —Offers fundamental resources like computing,
storage, and networking capabilities, using virtual servers such as Amazon EC2,
Google Compute Engine, and Microsoft Azure virtual machines
Platform as a service (PaaS) —Provides platforms to deploy custom applications to
the cloud, such as AWS Elastic Beanstalk, Google App Engine, and Heroku
Software as a service (SaaS) —Combines infrastructure and software running in
the cloud, including office applications like Amazon WorkSpaces, Google Apps
for Work, and Microsoft Office 365
The AWS product portfolio contains IaaS, PaaS, and SaaS. Let’s take a more concrete
look at what you can do with AWS.

1.2 What can you do with AWS?

You can run any application on AWS by using one or a combination of services. The
examples in this section will give you an idea of what you can do with AWS.

1.2.1 Hosting a web shop

John is CIO of a medium-sized e-commerce business. His goal is to provide his custom-
ers with a fast and reliable web shop. He decided to host the web shop on-premises,
and three years ago he rented servers in a data center. A web server handles requests
from customers, and a database stores product information and orders. John is evalu-
ating how his company can take advantage of AWS by running the same setup on AWS,
as shown in figure 1.2.

Licensed to Marocco Aurélien <[email protected]>

 What can you do with AWS? 5

User

Internet

On-premises server

Web
Database
server
Web server Database

Maintenance free Managed by you with updates, Figure 1.2 Running a web shop
monitoring, and so on on-premises vs. on AWS

John realized that other options are available to improve his setup on AWS with addi-
tional services:
The web shop consists of dynamic content (such as products and their prices)
and static content (such as the company logo). By splitting dynamic and static
content, John reduced the load for his web servers and improved performance
by delivering the static content over a content delivery network (CDN).
John uses maintenance-free services including a database, an object store, and a
DNS system on AWS. This frees him from managing these parts of the system,
decreases operational costs, and improves quality.
The application running the web shop can be installed on virtual servers. John
split the capacity of the old on-premises server into multiple smaller virtual serv-
ers at no extra cost. If one of these virtual servers fails, the load balancer will
send customer requests to the other virtual servers. This setup improves the web
shop’s reliability.
Figure 1.3 shows how John enhanced the web shop setup with AWS.
John started a proof-of-concept project and found that his web application can be
transferred to AWS and that services are available to help improve his setup.

1.2.2 Running a Java EE application in your private network

Maureen is a senior system architect in a global corporation. She wants to move parts
of the business applications to AWS when the company’s data-center contract expires
in a few months, to reduce costs and gain flexibility. She found that it’s possible to run
enterprise applications on AWS.

Licensed to Marocco Aurélien <[email protected]>

6 CHAPTER 1 What is Amazon Web Services?

Internet
User
Improve Improve
reliability performance

Dynamic Static

Load balancer DNS CDN

Decrease
maintenance Web server
costs

Object store

Database

Maintenance free Managed by you with updates,

monitoring, and so on

Figure 1.3 Running a web shop on AWS with CDN for better performance, a load balancer for
high availability, and a managed database to decrease maintenance costs

To do so, she defines a virtual network in the cloud and connects it to the corpo-
rate network through a virtual private network (VPN) connection. The company
can control access and protect mission-critical data by using subnets and control
traffic between them with access-control lists. Maureen controls traffic to the
internet using Network Address Translation (NAT) and firewalls. She installs
application servers on virtual machines (VMs) to run the Java EE application. Mau-
reen is also thinking about storing data in a SQL database service (such as Oracle
Database Enterprise Edition or Microsoft SQL Server EE). Figure 1.4 illustrates Mau-
reen’s architecture.
Maureen has managed to connect the on-premises data center with a private net-
work on AWS. Her team has already started to move the first enterprise application to
the cloud.

1.2.3 Meeting legal and business data archival requirements

Greg is responsible for the IT infrastructure of a small law office. His primary goal is to
store and archive all data in a reliable and durable way. He operates a file server to

Licensed to Marocco Aurélien <[email protected]>

 What can you do with AWS? 7

Virtual network
10.10.0.0/16

Private subnet
10.10.0.0/24
Internet
Internet
gateway NAT

Private subnet
VPN 10.10.1.0/24

VPN
Corporate network gateway Java EE server
10.20.0.0/16

Private subnet
10.10.2.0/24

SQL database

Figure 1.4 Running a Java EE application with enterprise networking on AWS

offer the possibility of sharing documents within the office. Storing all the data is a
challenge for him:
He needs to back up all files to prevent the loss of critical data. To do so, Greg
copies the data from the file server to another network-attached storage, so he
had to buy the hardware for the file server twice. The file server and the backup
server are located close together, so he is failing to meet disaster-recovery
requirements to recover from a fire or a break-in.
To meet legal and business data archival requirements, Greg needs to store data
for a long time. Storing data for 10 years or longer is tricky. Greg uses an expen-
sive archive solution to do so.
To save money and increase data security, Greg decided to use AWS. He transferred
data to a highly available object store. A storage gateway makes it unnecessary to buy
and operate network-attached storage and a backup on-premises. A virtual tape deck
takes over the task of archiving data for the required length of time. Figure 1.5
shows how Greg implemented this use case on AWS and compares it to the
on-premises solution.
Greg is fine with the new solution to store and archive data on AWS because he was
able to improve quality and he gained the possibility of scaling storage size.

Licensed to Marocco Aurélien <[email protected]>

8 CHAPTER 1 What is Amazon Web Services?

Local company network Local company network

User User User User User User

Backup Archive

NAS (backup) Network-attached Tape deck Storage gateway

storage (NAS)

Internet

Data storage in a single

location is a disaster risk.
Synchronize

With high-availability Archive

services, no backup
is required.
Object Virtual
store tape drive

Maintenance free Managed by you with updates,

monitoring, and so on

Figure 1.5 Backing up and archiving data on-premises and on AWS

1.2.4 Implementing a fault-tolerant system architecture

Alexa is a software engineer working for a fast-growing startup. She knows that Mur-
phy’s Law applies to IT infrastructure: anything that can go wrong, will go wrong. Alexa
is working hard to build a fault-tolerant system to prevent outages from ruining the
business. She knows that there are two type of services on AWS: fault-tolerant services
and services that can be used in a fault-tolerant way. Alexa builds a system like the one
shown in figure 1.6 with a fault-tolerant architecture. The database service is offered
with replication and failover handling. Alexa uses virtual servers acting as web servers.
These virtual servers aren’t fault tolerant by default. But Alexa uses a load balancer and
can launch multiple servers in different data centers to achieve fault tolerance.
So far, Alexa has protected the startup from major outages. Nevertheless, she and
her team are always planning for failure.
You now have a broad idea of what you can do with AWS. Generally speaking, you
can host any application on AWS. The next section explains the nine most important
benefits AWS has to offer.

Licensed to Marocco Aurélien <[email protected]>

 How you can benefit from using AWS 9

Data center A

Database
(master)

Web server

Internet
User Data center B
Load
balancer

Database
(standby)

Web server

Fault tolerant by default Highly available Fault tolerant usage possible

Figure 1.6 Building a fault-tolerant system on AWS

1.3 How you can benefit from using AWS

What’s the most important advantage of using AWS? Cost savings, you might say. But
saving money isn’t the only advantage. Let’s look at other ways you can benefit from
using AWS.

1.3.1 Innovative and fast-growing platform

In 2014, AWS announced more than 500 new services and features during its yearly
conference, re:Invent at Las Vegas. On top of that, new features and improvements
are released every week. You can transform these new services and features into inno-
vative solutions for your customers and thus achieve a competitive advantage.
The number of attendees to the re:Invent conference grew from 9,000 in 2013
to 13,500 in 2014.1 AWS counts more than 1 million businesses and government agen-
cies among its customers, and in its Q1 2014 results discussion, the company said it
will continue to hire more talent to grow even further.2 You can expect even more new
features and services in the coming years.

1
Greg Bensinger, “Amazon Conference Showcases Another Side of the Retailer’s Business,” Digits, Nov. 12, 2014,
http://mng.bz/hTBo.
2
“Amazon.com’s Management Discusses Q1 2014 Results - Earnings Call Transcript,” Seeking Alpha, April 24, 2014,
http://mng.bz/60qX.

Licensed to Marocco Aurélien <[email protected]>

10 CHAPTER 1 What is Amazon Web Services?

1.3.2 Services solve common problems

As you’ve learned, AWS is a platform of services. Common problems such as load bal-
ancing, queuing, sending email, and storing files are solved for you by services. You don’t
need to reinvent the wheel. It’s your job to pick the right services to build complex sys-
tems. Then you can let AWS manage those services while you focus on your customers.

1.3.3 Enabling automation

Because AWS has an API, you can automate everything: you can write code to create
networks, start virtual server clusters, or deploy a relational database. Automation
increases reliability and improves efficiency.
The more dependencies your system has, the more complex it gets. A human can
quickly lose perspective, whereas a computer can cope with graphs of any size. You
should concentrate on tasks a human is good at—describing a system—while the com-
puter figures out how to resolve all those dependencies to create the system. Setting
up an environment in the cloud based on your blueprints can be automated with the
help of infrastructure as code, covered in chapter 4.

1.3.4 Flexible capacity (scalability)

Flexible capacity frees you from planning. You can scale from one server to thousands
of servers. Your storage can grow from gigabytes to petabytes. You no longer need to
predict your future capacity needs for the coming months and years.
If you run a web shop, you have seasonal traffic patterns, as shown in figure 1.7.
Think about day versus night, and weekday versus weekend or holiday. Wouldn’t it be
nice if you could add capacity when traffic grows and remove capacity when traffic
shrinks? That’s exactly what flexible capacity is about. You can start new servers within
minutes and throw them away a few hours after that.
The cloud has almost no capacity constraints. You no longer need to think about
rack space, switches, and power supplies—you can add as many servers as you like. If
your data volume grows, you can always add new storage capacity.
System load

System load

System load

6am 12am 6pm Monday Thursday Sunday January December

Figure 1.7 Seasonal traffic patterns for a web shop

Licensed to Marocco Aurélien <[email protected]>

 How you can benefit from using AWS 11

Flexible capacity also means you can shut down unused systems. In one of our last proj-
ects, the test environment only ran from 7:00 a.m. to 8:00 p.m. on weekdays, allowing
us to save 60%.

1.3.5 Built for failure (reliability)

Most AWS services are fault-tolerant or highly available. If you use those services, you
get reliability for free. AWS supports you as you build systems in a reliable way. It pro-
vides everything you need to create your own fault-tolerant systems.

1.3.6 Reducing time to market

In AWS, you request a new virtual server, and a few minutes later that virtual server is
booted and ready to use. The same is true with any other AWS service available. You
can use them all on demand. This allows you to adapt your infrastructure to new
requirements very quickly.
Your development process will be faster because of the shorter feedback loops. You
can eliminate constraints such as the number of test environments available; if you
need one more test environment, you can create it for a few hours.

1.3.7 Benefiting from economies of scale

At the time of writing, the charges for using AWS have been reduced 42 times since 2008:
In December 2014, charges for outbound data transfer were lowered by up to 43%.
In November 2014, charges for using the search service were lowered by 50%.
In March 2014, charges for using a virtual server were lowered by up to 40%.
As of December 2014, AWS operated 1.4 million servers. All processes related to oper-
ations must be optimized to operate at that scale. The bigger AWS gets, the lower the
prices will be.

1.3.8 Worldwide
You can deploy your applications as close to your customers as possible. AWS has data
centers in the following locations:
United States (northern Virginia, northern California, Oregon)
Europe (Germany, Ireland)
Asia (Japan, Singapore)
Australia
South America (Brazil)
With AWS, you can run your business all over the world.

1.3.9 Professional partner

AWS is compliant with the following:
ISO 27001—A worldwide information security standard certified by an indepen-
dent and accredited certification body

Licensed to Marocco Aurélien <[email protected]>

12 CHAPTER 1 What is Amazon Web Services?

FedRAMP & DoD CSM —Ensures secure cloud computing for the U.S. Federal
Government and the U.S. Department of Defense
PCI DSS Level 1 —A data security standard (DSS) for the payment card industry
(PCI ) to protect cardholders data
ISO 9001 —A standardized quality management approach used worldwide and
certified by an independent and accredited certification body
If you’re still not convinced that AWS is a professional partner, you should know that
Airbnb, Amazon, Intuit, NASA, Nasdaq, Netflix, SoundCloud, and many more are run-
ning serious workloads on AWS.
The cost benefit is elaborated in more detail in the next section.

1.4 How much does it cost?

A bill from AWS is similar to an electric bill. Services are billed based on usage. You pay
for the hours a virtual server was running, the used storage from the object store (in
gigabytes), or the number of running load balancers. Services are invoiced on a
monthly basis. The pricing for each service is publicly available; if you want to calcu-
late the monthly cost of a planned setup, you can use the AWS Simple Monthly Calcu-
lator (http://aws.amazon.com/calculator).

1.4.1 Free Tier

You can use some AWS services for free during the first 12 months after you sign up.
The idea behind the Free Tier is to enable you to experiment with AWS and get some
experience. Here is what’s included in the Free Tier:
750 hours (roughly a month) of a small virtual server running Linux or Win-
dows. This means you can run one virtual server the whole month or you can
run 750 virtual servers for one hour.
750 hours (or roughly a month) of a load balancer.
Object store with 5 GB of storage.
Small database with 20 GB of storage, including backup.
If you exceed the limits of the Free Tier, you start paying for the resources you con-
sume without further notice. You’ll receive a bill at the end of the month. We’ll show
you how to monitor your costs before you begin using AWS. If your Free Tier ends
after one year, you pay for all resources you use.
You get some additional benefits, as detailed at http://aws.amazon.com/free. This
book will use the Free Tier as much as possible and will clearly state when additional
resources are required that aren’t covered by the Free Tier.

1.4.2 Billing example

As mentioned earlier, you can be billed in several ways:
Based on hours of usage —If you use a server for 61 minutes, that’s usually counted
as 2 hours.

Licensed to Marocco Aurélien <[email protected]>

 How much does it cost? 13

Based on traffic —Traffic can be measured in gigabytes or in number of requests.

Based on storage usage —Usage can be either provisioned capacity (for example, 50 GB
volume no matter how much you use) or real usage (such as 2.3 GB used).
Remember the web shop example from section 1.2? Figure 1.8 shows the web shop
and adds information about how each part is billed.

Internet
User

Dynamic Static

Load balancer DNS CDN

Web server

Object
storage

Database

Billed by hours of usage Billed by traffic Billed by storage usage

Figure 1.8 Web shop billing example

Let’s assume your web shop started successfully in January, and you decided to run a
marketing campaign to increase sales for the next month. Lucky you: you were able to
increase the number of visitors of your web shop fivefold in February. As you already
know, you have to pay for AWS based on usage. Table 1.1 shows your bills for January
and February. The number of visitors increased from 100,000 to 500,000, and your
monthly bill increased from 142.37 USD to 538.09 USD, which is a 3.7-fold increase.
Because your web shop had to handle more traffic, you had to pay more for services,
such as the CDN, the web servers, and the database. Other services, like the storage of
static files, didn’t experience more usage, so the price stayed the same.
With AWS, you can achieve a linear relationship between traffic and costs. And
other opportunities await you with this pricing model.

Licensed to Marocco Aurélien <[email protected]>

14 CHAPTER 1 What is Amazon Web Services?

Table 1.1 How an AWS bill changes if the number of web shop visitors increases

Service January usage February usage February charge Increase

Visits to website 100,000 500,000

CDN 26 M requests + 131 M requests + 113.31 USD 90.64 USD

25 GB traffic 125 GB traffic

Static files 50 GB used 50 GB used 1.50 USD 0.00 USD

storage storage

Load balancer 748 hours + 748 hours + 20.30 USD 1.60 USD
50 GB traffic 250 GB traffic

Web servers 1 server = 748 4 servers = 2,992 204.96 USD 153.72 USD
hours hours

Database (748 Small server + Large server + 170.66 USD 128.10 USD
hours) 20 GB storage 20 GB storage

Traffic (outgoing 51 GB 255 GB 22.86 USD 18.46 USD

traffic to internet)

DNS 2 M requests 10 M requests 4.50 USD 3.20 USD

Total cost 538.09 USD 395.72 USD

1.4.3 Pay-per-use opportunities

The AWS pay-per-use pricing model creates new opportunities. You no longer need to
make upfront investments in infrastructure. You can start servers on demand and only
pay per hour of usage; and you can stop using those servers whenever you like and no
longer have to pay for them. You don’t need to make an upfront commitment regard-
ing how much storage you’ll use.
A big server costs exactly as much as two smaller ones with the same capacity. Thus
you can divide your systems into smaller parts, because the cost is the same. This
makes fault tolerance affordable not only for big companies but also for smaller
budgets.

1.5 Comparing alternatives

AWS isn’t the only cloud computing provider. Microsoft and Google have cloud offer-
ings as well.
OpenStack is different because it’s open source and developed by more than 200
companies including IBM, HP, and Rackspace. Each of these companies uses Open-
Stack to operate its own cloud offerings, sometimes with closed source add-ons. You
could run your own cloud based on OpenStack, but you would lose most of the bene-
fits outlined in section 1.3.
Comparing cloud providers isn’t easy, because open standards are mostly missing.
Functionality like virtual networks and message queuing are realized differently. If you
know what features you need, you can compare the details and make your decision.

Licensed to Marocco Aurélien <[email protected]>

 Comparing alternatives 15

Otherwise, AWS is your best bet because the chances are highest that you’ll find a solu-
tion for your problem.
Following are some common features of cloud providers:
Virtual servers (Linux and Windows)
Object store
Load balancer
Message queuing
Graphical user interface
Command-line interface
The more interesting question is, how do cloud providers differ? Table 1.2 compares
AWS, Azure, Google Cloud Platform, and OpenStack.

Table 1.2 Differences between AWS, Microsoft Azure, Google Cloud Platform, and OpenStack

Google Cloud
AWS Azure OpenStack
Platform

Number of services Most Many Enough Few

Number of locations 9 13 3 Yes (depends on the

(multiple data cen- OpenStack provider)
ters per location)

Compliance Common standards Common standards Common standards Yes (depends on the
(ISO 27001, HIPAA, (ISO 27001, HIPAA, (ISO 27001, HIPAA, OpenStack provider)
FedRAMP, SOC), IT FedRAMP, SOC), ISO FedRAMP, SOC)
Grundschutz (Ger- 27018 (cloud pri-
many), G-Cloud (UK) vacy), G-Cloud (UK)

SDK languages Android, Browsers Android, iOS, Java, Java, Browsers -

(JavaScript), iOS, .NET, Node.js (JavaS- (JavaScript), .NET,
Java, .NET, Node.js cript), PHP, Python, PHP, Python
(JavaScript), PHP, Ruby
Python, Ruby, Go

Integration into Medium, not linked High, linked to the High, linked to the -
development to specific ecosys- Microsoft ecosys- Google ecosystem
process tems tem (for example, (for example,
.NET development) Android)

Block-level storage Yes Yes (can be used by No Yes (can be used by

(attached via net- multiple virtual serv- multiple virtual serv-
work) ers simultaneously) ers simultaneously)

Relational Yes (MySQL, Post- Yes (Azure SQL Data- Yes (MySQL) Yes (depends on the
database greSQL, Oracle Data- base, Microsoft SQL OpenStack provider)
base, Microsoft SQL Server)
Server)

NoSQL database Yes (proprietary) Yes (proprietary) Yes (proprietary) No

DNS Yes No Yes No

Licensed to Marocco Aurélien <[email protected]>

16 CHAPTER 1 What is Amazon Web Services?

Table 1.2 Differences between AWS, Microsoft Azure, Google Cloud Platform, and OpenStack (continued)

Google Cloud
AWS Azure OpenStack
Platform

Virtual network Yes Yes No Yes

Pub/sub messag- Yes (proprietary, JMS Yes (proprietary) Yes (proprietary) No

ing library available)

Machine-learning Yes Yes Yes No

tools

Deployment tools Yes Yes Yes No

On-premises data- Yes Yes Yes No

center integration

In our opinion, AWS is the most mature cloud platform available at the moment.

1.6 Exploring AWS services

Hardware for computing, storing, and networking is the foundation of the AWS cloud.
AWS runs software services on top of the hardware to provide the cloud, as shown in
figure 1.9. A web interface, the API, acts as an interface between AWS services and your
applications.
You can manage services by sending requests to the API manually via a GUI or pro-
grammatically via a SDK. To do so, you can use a tool like the Management Console, a
web-based user interface, or a command-line tool. Virtual servers have a peculiarity:
you can connect to virtual servers through SSH, for example, and gain administrator

Manage
services
Services
API
Administrator Compute: Virtual server
App: Queues, search
Enterprise: Directory service, mail
Deployment: Access rights, monitoring
Storage: Object store, archiving
Database: Relational, NoSQL
Networking: DNS, virtual network

Software
Hardware

Compute Storage
Network

Figure 1.9 The AWS cloud is composed of hardware and software services accessible via an API.

Licensed to Marocco Aurélien <[email protected]>

 Exploring AWS services 17

Administrator

Install and configure

software remotely

Manage
services

API
Services

Virtual
server

NoSQL Sending Static file

database email storage

Figure 1.10 Managing a custom application running on a virtual server and dependent services

access. This means you can install any software you like on a virtual server. Other ser-
vices, like the NoSQL database service, offer their features through an API and hide
everything that’s going on behind the scenes. Figure 1.10 shows an administrator
installing a custom PHP web application on a virtual server and managing dependent
services such as a NoSQL database used by the PHP web application.
Users send HTTP requests to a virtual server. A web server is installed on this virtual
server along with a custom PHP web application. The web application needs to talk to
AWS services in order to answer HTTP requests from users. For example, the web
application needs to query data from a NoSQL database, store static files, and send
email. Communication between the web application and AWS services is handled by
the API, as figure 1.11 shows.
The number of different services available can be scary at the outset. The following
categorization of AWS services will help you to find your way through the jungle:
Compute services offer computing power and memory. You can start virtual serv-
ers and use them to run your applications.
App services offer solutions for common use cases like message queues, topics,
and searching large amounts of data to integrate into your applications.

Licensed to Marocco Aurélien <[email protected]>

18 CHAPTER 1 What is Amazon Web Services?

Enterprise services offer independent solutions such as mail servers and directory
services.
Deployment and administration services work on top of the services mentioned so
far. They help you grant and revoke access to cloud resources, monitor your vir-
tual servers, and deploy applications.
Storage is needed to collect, persist, and archive data. AWS offers different stor-
age options: an object store or a network-attached storage solution for use with
virtual servers.
Database storage has some advantages over simple storage solutions when you
need to manage structured data. AWS offers solutions for relational and NoSQL
databases.
Networking services are an elementary part of AWS. You can define private net-
works and use a well-integrated DNS.
Be aware that we cover only the most important categories and services here. Other
services are available, and you can also run your own applications on AWS.
Now that we’ve looked at AWS services in detail, it’s time for you to learn how to
interact with those services.

Users

HTTP request

API Services

Virtual
server

NoSQL Sending Static file

database email storage

Figure 1.11 Handling an HTTP request with a custom web application using additional
AWS services

Licensed to Marocco Aurélien <[email protected]>

 Interacting with AWS 19

1.7 Interacting with AWS

When you interact with AWS to configure or use services, you make calls to the API.
The API is the entry point to AWS, as figure 1.12 demonstrates.
Next, we’ll give you an overview of the tools available to make calls to the AWS API.
You can compare the ability of these tools to automate your daily tasks.

1.7.1 Management Console

You can use the web-based Management Console to interact with AWS. You can manu-
ally control AWS with this convenient GUI, which runs in every modern web browser
(Chrome, Firefox, Safari 5, IE 9); see figure 1.13.
If you’re experimenting with AWS, the Management Console is the best place to
start. It helps you to gain an overview of the different services and achieve success
quickly. The Management Console is also a good way to set up a cloud infrastructure
for development and testing.

1.7.2 Command-line interface

You can start a virtual server, create storage, and send email from the command line.
With the command-line interface (CLI), you can control everything on AWS; see fig-
ure 1.14.

Manual

Command-
Web-based
line interface
management
Console

Automation
SDKs for Java,
Python, JavaScript,...

API

Blueprints
Services

Figure 1.12 Tools to interact with the AWS API

Licensed to Marocco Aurélien <[email protected]>

20 CHAPTER 1 What is Amazon Web Services?

Figure 1.13 Management Console

Figure 1.14 Command-line interface

Licensed to Marocco Aurélien <[email protected]>

 Interacting with AWS 21

The CLI is typically used to automate tasks on AWS. If you want to automate parts of
your infrastructure with the help of a continuous integration server like Jenkins, the
CLI is the right tool for the job. The CLI offers a convenient way to access the API and
combine multiple calls into a script.
You can even begin to automate your infrastructure with scripts by chaining multi-
ple CLI calls together. The CLI is available for Windows, Mac, and Linux, and there’s
also a PowerShell version available.

1.7.3 SDKs
Sometimes you need to call AWS from within your application. With SDKs, you can use
your favorite programming language to integrate AWS into your application logic. AWS
provides SDKs for the following:
Android Node.js (JavaScript)
Browsers (JavaScript) PHP
iOS Python
Java Ruby
.NET Go

SDKs are typically used to integrate AWS services into applications. If you’re doing soft-
ware development and want to integrate an AWS service like a NoSQL database or a
push-notification service, an SDK is the right choice for the job. Some services, such as
queues and topics, must be used with an SDK in your application.

1.7.4 Blueprints
A blueprint is a description of your system containing all services and dependencies. The
blueprint doesn’t say anything about the necessary steps or the order to achieve the
described system. Figure 1.15 shows how a blueprint is transferred into a running system.

DNS CDN
{
infrastructure: {
loadbalancer: {
server: { ... }
}, Tool
cdn: { ... },
database: { ... },
Load balancer Static files
dns: { ... },
static: { ... }
}
}

Figure 1.15 Infrastructure

Web servers Database automation with blueprints

Licensed to Marocco Aurélien <[email protected]>

22 CHAPTER 1 What is Amazon Web Services?

Consider using blueprints if you have to control many or complex environments.

Blueprints will help you to automate the configuration of your infrastructure in the
cloud. You can use blueprints to set up virtual networks and launch different servers
into that network, for example.
A blueprint removes much of the burden from you because you no longer need to
worry about dependencies during system creation—the blueprint automates the
entire process. You’ll learn more about automating your infrastructure in chapter 4.
It’s time to get started creating your AWS account and exploring AWS practice after
all that theory.

1.8 Creating an AWS account

Before you can start using AWS, you need to create an account. An AWS account is a
basket for all the resources you own. You can attach multiple users to an account if
multiple humans need access to the account; by default, your account will have one
root user. To create an account, you need the following:
A telephone number to validate your identity
A credit card to pay your bills

Using an old account?

You can use your existing AWS account while working on the examples in this book. In
this case, your usage may not be covered by the Free Tier, and you may have to pay for
your usage.
Also, if you created your existing AWS account before December 4, 2013, you should create
a new one: there are legacy issues that may cause trouble when you try our examples.

1.8.1 Signing up
The sign-up process consists of five steps:
1 Provide your login credentials.
2 Provide your contact information.
3 Provide your payment details.
4 Verify your identity.
5 Choose your support plan.
Point your favorite modern web browser to https://aws.amazon.com, and click the
Create a Free Account / Create an AWS Account button.
1. PROVIDING YOUR LOGIN CREDENTIALS
The Sign Up page, shown in figure 1.16, gives you two choices. You can either create
an account using your Amazon.com account or create an account from scratch. If you
create the account from scratch, follow along. Otherwise, skip to step 5.
Fill in your email address, and select I Am a New User. Go on to the next step to cre-
ate your login credentials. We advise you to choose a strong password to prevent misuse

Licensed to Marocco Aurélien <[email protected]>

 Creating an AWS account 23

Figure 1.16 Creating an

AWS account: Sign Up page

of your account. We suggest a password with 16 characters, numbers, and symbols. If

someone gets access to your account, they can destroy your systems or steal your data.
2. PROVIDING YOUR CONTACT INFORMATION
The next step, as shown in figure 1.17, is to provide your contact information. Fill in
all the required fields, and continue.

Figure 1.17 Creating an

AWS account: providing
your contact information

Licensed to Marocco Aurélien <[email protected]>

24 CHAPTER 1 What is Amazon Web Services?

Figure 1.18 Creating an AWS account: providing your payment details

3. PROVIDE YOUR PAYMENT DETAILS

Now the screen shown in figure 1.18 asks for your payment information. AWS supports
MasterCard and Visa. You can set your preferred payment currency later, if you don’t
want to pay your bills in USD; supported currencies are EUR, GBP, CHF, AUD, and some
others.
4. VERIFYING YOUR IDENTITY
The next step is to verify your identity. Figure 1.19 shows the first step of the process.

Licensed to Marocco Aurélien <[email protected]>

 Creating an AWS account 25

Figure 1.19 Creating an AWS account: verifying your identity (1 of 2)

After you complete the first part, you’ll receive a call from AWS. A robot voice will ask
you for your PIN, which will be like the one shown in figure 1.20. Your identity will be
verified, and you can continue with the last step.

Figure 1.20 Creating an AWS account: verifying your identity (2 of 2)

Licensed to Marocco Aurélien <[email protected]>

26 CHAPTER 1 What is Amazon Web Services?

5. CHOOSING YOUR SUPPORT PLAN

The last step is to choose a support plan; see figure 1.21. In this case, select the Basic
plan, which is free. If you later create an AWS account for your business, we recom-
mend the Business support plan. You can even switch support plans later.
High five! You’re done. Now you can log in to your account with the AWS Manage-
ment Console.

Figure 1.21 Creating an AWS account: choosing your support plan

1.8.2 Signing In
You have an AWS account and are ready to sign in to the AWS Management Console at
https://console.aws.amazon.com. As mentioned earlier, the Management Console is
a web-based tool you can use to control AWS resources. The Management Console

Licensed to Marocco Aurélien <[email protected]>

 Creating an AWS account 27

Figure 1.22 Sign in to the Management Console.

uses the AWS API to make most of the functionality available to you. Figure 1.22 shows
the Sign In page.
Enter your login credentials and click Sign In Using Our Secure Server to see the
Management Console, shown in figure 1.23.

Figure 1.23 AWS Management Console

Licensed to Marocco Aurélien <[email protected]>

28 CHAPTER 1 What is Amazon Web Services?

The most important part is the navigation bar at the top; see figure 1.24. It consists of
six sections:
AWS —Gives you a fast overview of all resources in your account.
Services —Provides access to all AWS services.
Custom section (Edit) —Click Edit and drag-and-drop important services here to
personalize the navigation bar.
Your name —Lets you access billing information and your account, and also lets
you sign out.
Your region —Lets you choose your region. You’ll learn about regions in section
3.5. You don’t need to change anything here now.
Support —Gives you access to forums, documentation, and a ticket system.

Resource Quick access to services Region Help

overview (customizable) selector section
Jump to Account and
a service billing

Figure 1.24 AWS Management Console navigation bar

Next, you’ll create a key pair so you can connect to your virtual servers.

1.8.3 Creating a key pair

To access a virtual server in AWS, you need a key pair consisting of a private key and a
public key. The public key will be uploaded to AWS and inserted into the virtual server.
The private key is yours; it’s like your password, but much more secure. Protect your
private key as if it’s a password. It’s your secret, so don’t lose it—you can’t retrieve it.
To access a Linux server, you use the SSH protocol; you’ll authenticate with the
help of your key pair instead of a password during login. You access a Windows server
via Remote Desktop Protocol (RDP); you’ll need your key pair to decrypt the adminis-
trator password before you can log in.
The following steps will guide you to the dashboard of the EC2 service, which offers
virtual servers, and where you can obtain a key pair:
1 Open the AWS Management Console at https://console.aws.amazon.com.
2 Click Services in the navigation bar, find the EC2 service, and click it.
3 Your browser shows the EC2 Management Console.
The EC2 Management Console, shown in figure 1.25, is split into three columns. The first
column is the EC2 navigation bar; because EC2 is one of the oldest services, it has many

Licensed to Marocco Aurélien <[email protected]>

 Creating an AWS account 29

Figure 1.25 EC2 Management Console

features that you can access via the navigation bar. The second column gives you a brief
overview of all your EC2 resources. The third column provides additional information.
Follow these steps to create a new key pair:
1 Click Key Pairs in the navigation bar under Network & Security.
2 Click the Create Key Pair button on the page shown in figure 1.26.
3 Name the Key Pair mykey. If you choose another name, you must replace the
name in all the following examples!
During key-pair creation, you downloaded a file called mykey.pem. You must now pre-
pare that key for future use. Depending on your operating system, you may need to do
things differently, so please read the section that fits your OS.

Using your own key pair

It’s also possible to upload the public key part from an existing key pair to AWS. Doing
so has two advantages:
You can reuse an existing key pair.
You can be sure that only you know the private key part of the key pair. If you use
the Create Key Pair button, AWS knows (at least briefly) your private key.
We decided against that approach in this case because it’s less convenient to imple-
ment in a book.

Licensed to Marocco Aurélien <[email protected]>

30 CHAPTER 1 What is Amazon Web Services?

Figure 1.26 EC2 Management Console key pairs

LINUX AND MAC OS X

The only thing you need to do is change the access rights of mykey.pem so that only
you can read the file. To do so, run chmod 400 mykey.pem in the terminal. You’ll learn
about how to use your key when you need to log in to a virtual server for the first time
in this book.
WINDOWS
Windows doesn’t ship a SSH client, so you need to download the PuTTY installer for
Windows from http://mng.bz/A1bY and install PuTTY. PuTTY comes with a tool
called PuTTYgen that can convert the mykey.pem file into a mykey.ppk file, which
you’ll need:
1 Run the application PuTTYgen. The screen shown in figure 1.27 opens.
2 Select SSH-2 RSA under Type of Key to Generate.
3 Click Load.
4 Because PuTTYgen displays only *.pkk files, you need to switch the file exten-
sion of the File Name field to All Files.
5 Select the mykey.pem file, and click Open.
6 Confirm the dialog box.
7 Change Key Comment to mykey.
8 Click Save Private Key. Ignore the warning about saving the key without a
passphrase.
Your .pem file is now converted to the .pkk format needed by PuTTY. You’ll learn how
to use your key when you need to log in to a virtual server for the first time in this book.

Licensed to Marocco Aurélien <[email protected]>

 Summary 31

Figure 1.27 PuTTYgen allows you to convert the downloaded .pem file into the .pkk
file format needed by PuTTY.

1.8.4 Creating a billing alarm

Before you use your AWS account in the next chapter, we advise you to create a billing
alarm. If you exceed the Free Tier, an email is sent to you. The book warns you when-
ever an example isn’t covered by the Free Tier. Please make sure that you carefully fol-
low the cleanup steps after each example. To make sure you haven’t missed something
during cleanup, please create a billing alarm as advised by AWS: http://mng.bz/M7Sj.

1.9 Summary
Amazon Web Services (AWS) is a platform of web services offering solutions for
computing, storing, and networking that work well together.
Cost savings aren’t the only benefit of using AWS. You’ll also profit from an
innovative and fast-growing platform with flexible capacity, fault-tolerant ser-
vices, and a worldwide infrastructure.
Any use case can be implemented on AWS, whether it’s a widely used web appli-
cation or a specialized enterprise application with an advanced networking
setup.

Licensed to Marocco Aurélien <[email protected]>

32 CHAPTER 1 What is Amazon Web Services?

You can interact with AWS in many different ways. You can control the different ser-
vices by using the web-based GUI; use code to manage AWS programmatically from
the command line or SDKs; or use blueprints to set up, modify, or delete your infra-
structure on AWS.
Pay-per-use is the pricing model for AWS services. Computing power, storage,
and networking services are billed similarly to electricity.
Creating an AWS account is easy. Now you know how to set up a key pair so you
can log in to virtual servers for later use.

Licensed to Marocco Aurélien <[email protected]>

 Summary 33

Physical data centers require lots of equipment and

take time and resources to manage. If you need a data
center, but don’t want to build your own, Amazon Web
Services may be your solution. Whether you’re analyz-
ing real-time data, building software as a service, or
running an e-commerce site, AWS offers you a reliable
cloud-based platform with services that scale.
Amazon Web Services in Action introduces you to
computing, storing, and networking in the AWS cloud.
The book will teach you about the most important ser-
vices on AWS. You will also learn about best practices
regarding security, high availability and scalability.
You'll start with a broad overview of cloud computing and AWS and learn how to spin-
up servers manually and from the command line. You'll learn how to automate your
infrastructure by programmatically calling the AWS API to control every part of AWS.
You will be introduced to the concept of Infrastructure as Code with the help of AWS
CloudFormation. You will learn about different approaches to deploy applications on
AWS. You'll also learn how to secure your infrastructure by isolating networks, control-
ling traffic and managing access to AWS resources. Next, you'll learn options and tech-
niques for storing your data. You will experience how to integrate AWS services into
your own applications by the use of SDKs. Finally, this book teaches you how to design
for high availability, fault tolerance, and scalability.

What's inside
Overview of AWS cloud concepts and best practices
Manage servers on EC2 for cost-effectiveness
Infrastructure automation with Infrastructure as Code (AWS CloudFormation)
Deploy applications on AWS
Store data on AWS: SQL, NoSQL, object storage and block storage
Integrate Amazon's pre-built services
Architect highly available and fault tolerant systems

Written for developers and DevOps engineers moving distributed applications to the
AWS platform.

Licensed to Marocco Aurélien <[email protected]>

 Docker in Practice

A docker was a labourer who was responsible for load goods of all sizes and
shapes to a ship. Docker is also the name of a tool, helping you to deliver your
applications to all kinds of machines. By using standardization, Docker allows
you to automate packaging and deploying your applications. A nice side-effect is
that this automation works on your local development machine as well as on
cloud infrastructure. Docker in Practice helps you to get started with Docker. The
first chapter, “Discovering Docker,” explains the key concepts behind Docker.

Licensed to Marocco Aurélien <[email protected]>

 Chapter 1 from Docker in Practice by Ian Miell
and Aidan Hobson Sayers.

Discovering Docker

This chapter covers

What Docker is
The uses of Docker and how it can save you
time and money
The differences between containers and images
Docker’s layering feature
Building and running a to-do application using
Docker

Docker is a platform that allows you to “build, ship, and run any app, anywhere.” It
has come a long way in an incredibly short time and is now considered a standard
way of solving one of the costliest aspects of software: deployment.
Before Docker came along, the development pipeline typically consisted of
combinations of various technologies for managing the movement of software,
such as virtual machines, configuration management tools, different package man-
agement systems, and complex webs of library dependencies. All these tools

35

Licensed to Marocco Aurélien <[email protected]>

36 CHAPTER 1 Discovering Docker

needed to be managed and maintained by specialist engineers, and most had their
own unique ways of being configured.
Docker has changed all of this, allowing different engineers involved in this pro-
cess to effectively speak one language, making working together a breeze. Everything
goes through a common pipeline to a single output that can be used on any target—
there’s no need to continue maintaining a bewildering array of tool configurations, as
shown in figure 1.1.
At the same time, there’s no need to throw away your existing software stack if it
works for you—you can package it up in a Docker container as-is for others to con-
sume. As a bonus, you can see how these containers were built, so if you need to dig
into the details, you can.
This book is aimed at intermediate developers with some knowledge of Docker. If
you’re OK with the basics, feel free to skip to the later chapters. The goal of this book
is to expose the real-world challenges that Docker brings and show how they can be
overcome. But first we’re going to provide a quick refresher on Docker itself. If you
want a more thorough treatment of Docker’s basics, take a look at Docker in Action by
Jeff Nickoloff (Manning Publications, 2016).
In chapter 2 you’ll be introduced to Docker’s architecture more deeply with the
aid of some techniques that demonstrate its power. In this chapter you’re going to
learn what Docker is, see why it’s important, and start using it.

Life before Docker Life with Docker

Configuration Code Configuration Code

Vagrant Jenkins Chef Docker build

Release
Development Test Live Development Testing
to live

Inputs to the system requiring manual maintenance—fewer

inputs here mean less of a maintenance burden.

Tools that use the inputs to create environments

for software development

Stages of software development requiring

an environment to run in

Figure 1.1 How Docker has eased the tool maintenance burden

Licensed to Marocco Aurélien <[email protected]>

 The what and why of Docker 37

1.1 The what and why of Docker

Before we get our hands dirty, we’re going to discuss Docker a little so that you under-
stand its context, where the name “Docker” came from, and why we’re using it at all!

1.1.1 What is Docker?

To understand what Docker is, it’s easier to start with a metaphor than a technical expla-
nation, and the Docker metaphor is a powerful one. A docker was a labourer who
moved commercial goods into and out of ships when they docked at ports. There were
boxes and items of differing sizes and shapes, and experienced dockers were prized for
their ability to fit goods into ships by hand in cost-effective ways (see figure 1.2). Hiring
people to move stuff around wasn’t cheap, but there was no alternative.

Ship on which the

items were loaded

Teams of dockers
required to load
differently shaped
items onto ship

Ship can be designed to carry, load, and unload

predictably shaped items more efficiently.

Only one docker needed to Single container with different items in it. It
operate machines designed doesn't matter to the carrier what's inside the
to move containers container. The carrier can be loaded up elsewhere,
reducing the bottleneck of loading at port.

Figure 1.2 Shipping before and after standardized containers

Licensed to Marocco Aurélien <[email protected]>

38 CHAPTER 1 Discovering Docker

This should sound familiar to anyone working in software. Much time and intellectual
energy is spent getting metaphorically odd-shaped software into different sized meta-
phorical ships full of other odd-shaped software, so they can be sold to users or busi-
nesses elsewhere.
Figure 1.3 shows how time and money can be saved with the Docker concept.
Before Docker, deploying software to different environments required significant
effort. Even if you weren’t hand-running scripts to provision software on different
machines (and plenty of people still do exactly that), you’d still have to wrestle with con-
figuration management tools that manage state on what are increasingly fast-moving
environments starved of resources. Even when these efforts were encapsulated in VMs,
a lot of time was spent managing the deployment of these VMs, waiting for them to boot,
and managing the overhead of resource use they created.
With Docker, the configuration effort is separated from the resource management,
and the deployment effort is trivial: run docker run, and the environment’s image is
pulled down and ready to run, consuming fewer resources and contained so that it
doesn’t interfere with other environments.
You don’t need to worry about whether your container is going to be shipped to a
RedHat machine, an Ubuntu machine, or a CentOS VM image; as long as it has
Docker on it, it’ll be good to go.

Life before Docker

Install, configure, Install, configure, Install, configure,

and maintain complex and maintain complex and maintain complex
application application application
Three times the
effort to manage
deployment Dev laptop Test server Live server

Life with Docker

Install, configure,
docker run
and maintain complex Dev laptop
application

docker run
A single effort Docker image Test server
to manage
deployment
docker run
Live server

Figure 1.3 Software delivery before and after Docker

Licensed to Marocco Aurélien <[email protected]>

 The what and why of Docker 39

1.1.2 What is Docker good for?

Some crucial practical questions arise: why would you use Docker, and for what? The short
answer to the “why” is that for a modicum of effort, Docker can save your business a lot
of money quickly. Some of these ways (and by no means all) are discussed in the following
subsections. We’ve seen all of these benefits first-hand in real working contexts.
REPLACES VIRTUAL MACHINES (VMS)
Docker can be used to replace VMs in many situations. If you only care about the
application, not the operating system, Docker can replace the VM, and you can leave
worrying about the OS to someone else. Not only is Docker quicker than a VM to spin
up, it’s more lightweight to move around, and due to its layered filesystem, it’s much
easier and quicker to share changes with others. It’s also firmly rooted in the com-
mand line and is eminently scriptable.
PROTOTYPING SOFTWARE
If you want to quickly experiment with software without either disrupting your exist-
ing setup or going through the hassle of provisioning a VM, Docker can give you a
sandbox environment in milliseconds. The liberating effect of this is difficult to grasp
until you experience it for yourself.
PACKAGING SOFTWARE
Because a Docker image has effectively no dependencies for a Linux user, it’s a great
way to package software. You can build your image and be sure that it can run on any
modern Linux machine—think Java, without the need for a JVM.
ENABLING A MICROSERVICES ARCHITECTURE
Docker facilitates the decomposition of a complex system to a series of composable
parts, which allows you to reason about your services in a more discrete way. This can
allow you to restructure your software to make its parts more manageable and plugga-
ble without affecting the whole.
MODELLING NETWORKS
Because you can spin up hundreds (even thousands) of isolated containers on one
machine, modelling a network is a breeze. This can be great for testing real-world sce-
narios without breaking the bank.
ENABLING FULL-STACK PRODUCTIVITY WHEN OFFLINE
Because you can bundle all the parts of your system into Docker containers, you can
orchestrate these to run on your laptop and work on the move, even when offline.
REDUCING DEBUGGING OVERHEAD
Complex negotiations between different teams about software delivered is commonplace
within the industry. We’ve personally experienced countless discussions about broken
libraries; problematic dependencies; updates applied wrongly, or in the wrong order, or
even not performed at all; unreproducible bugs, and so on. It’s likely you have too. Docker
allows you to state clearly (even in script form) the steps for debugging a problem on a sys-
tem with known properties, making bug and environment reproduction a much simpler
affair, and one normally separated from the host environment provided.

Licensed to Marocco Aurélien <[email protected]>

40 CHAPTER 1 Discovering Docker

DOCUMENTING SOFTWARE DEPENDENCIES AND TOUCHPOINTS

By building your images in a structured way, ready to be moved to different environ-
ments, Docker forces you to document your software dependencies explicitly from a
base starting point. Even if you decide not to use Docker everywhere, this need to doc-
ument can help you install your software in other places.
ENABLING CONTINUOUS DELIVERY
Continuous delivery (CD) is a paradigm for software delivery based on a pipeline that
rebuilds the system on every change and then delivers to production (or “live”)
through an automated (or partly automated) process.
Because you can control the build environment’s state more exactly, Docker builds
are more reproducible and replicable than traditional software building methods.
This makes implementing CD much easier. Standard CD techniques such as
Blue/Green deployment (where “live” and “last” deployments are maintained on live)
and Phoenix Deployment (where whole systems are rebuilt on each release) are made
trivial by implementing a reproducible Docker-centric build process.
Now you know a bit about how Docker can help you. Before we dive into a real
example, let’s go over a couple of core concepts.

1.1.3 Key concepts

In this section we’re going to cover some key Docker concepts, which are illustrated in
figure 1.4.

Containers: A container is a
running instance of an image.
You can have multiple containers
running from the same image.

Docker host machine

Stored Running
on disk processes

MyApplication
container (v1) 1
Images: An image is a
collection of filesystem
layers and some metadata. Debian layer: MyApplication
/bin container (v1) 2
Taken together, they can be
/boot
spun up as Docker …
containers. /tmp MyApplication
/var container (v1) 3
Layers: A layer is a MyApplication
collection of changes code layer
to files. The differences
between v1 and v2 of MyApplication MyApplication
MyApplication are v2 layer container (v2) 1
stored in this layer.

Figure 1.4 Core Docker concepts

Licensed to Marocco Aurélien <[email protected]>

 The what and why of Docker 41

It’s most useful to get the concepts of images, containers, and layers clear in your
mind before you start running Docker commands. In short, containers are running sys-
tems defined by images. These images are made up of one or more layers (or sets of
diffs) plus some metadata for Docker.
Let’s look at some of the core Docker commands. We’ll turn images into contain-
ers, change them, and add layers to new images that we’ll commit. Don’t worry if all of
this sounds confusing. By the end of the chapter it will all be much clearer!
KEY DOCKER COMMANDS
Docker’s central function is to build, ship, and run software in any location that has
Docker.
To the end user, Docker is a command-line program that you run. Like git (or
any source control tool), this program has subcommands that perform different
operations.
The principal Docker subcommands you’ll use on your host are listed in table 1.1.

Table 1.1 Docker subcommands

Command Purpose

docker build Build a Docker image.

docker run Run a Docker image as a container.

docker commit Commit a Docker container as an image.

docker tag Tag a Docker image.

IMAGES AND CONTAINERS

If you’re unfamiliar with Docker, this may be the first time you’ve come across the
words “container” and “image” in this context. They’re probably the most important
concepts in Docker, so it’s worth spending a bit of time to make sure the difference
is clear.
In figure 1.5 you’ll see an illustration of these concepts, with three containers
started up from one base image.
One way to look at images and containers is to see them as analogous to programs
and processes. In the same way a process can be seen as an application being exe-
cuted, a Docker container can be viewed as a Docker image in execution.
If you’re familiar with object-oriented principles, another way to look at images
and containers is to view images as classes and containers as objects. In the same way
that objects are concrete instantiations of classes, containers are instantiations of
images. You can create multiple containers from a single image, and they are all iso-
lated from one another in the same way objects are. Whatever you change in the
object, it won’t affect the class definition—they’re fundamentally different things.

Licensed to Marocco Aurélien <[email protected]>

42 CHAPTER 1 Discovering Docker

A Docker image consists of files and metadata.

Image files take up This is the base image for the containers below.
most of the space.
Because of the isolation The metadata has
each container provides, information on
they must have their environment variables,
own copy of any required Docker image: Ubuntu port mappings, volumes,
tools, including language Files: Metadata: and other details we'll
environments or libraries. /bin/bash Port mappings discuss later.
/bin/bunzip2 Environment variables
Containers run one process /bin/bzcat
on startup. When this process […]
completes, the container stops. /var/spool/rsyslog
This startup process can /var/tmp
spawn others.

Ubuntu container 1 Ubuntu container 2 Ubuntu container 3

Process: nodejs Process: mysql Process: apache

Diffs from Ubuntu image: Diffs from Ubuntu image: Diffs from Ubuntu image:
MODIFIED: /opt/app/nodejs.log DELETE: /etc/nologin ADDED: //var/log/apache/apache.log

Changes to files are stored Containers are created from images, inherit
within the container in a their filesystems, and use their metadata to
copy-on-write mechanism. determine their startup configuration.
The base image cannot be Containers are separate but can be
affected by a container. configured to communicate with
each other.

Figure 1.5 Docker images and containers

1.2 Building a Docker application

We’re going to get our hands dirty now by building a simple “to-do” application (todo-
app) image with Docker. In the process, you’ll see some key Docker features like
Dockerfiles, image re-use, port exposure, and build automation. Here’s what you’ll
learn in the next 10 minutes:
How to create a Docker image using a Dockerfile
How to tag a Docker image for easy reference
How to run your new Docker image
A to-do app is one that helps you keep track of things you want to get done. The app
we’ll build will store and display short strings of information that can be marked as
done, presented in a simple web interface.

Licensed to Marocco Aurélien <[email protected]>

 Building a Docker application 43

ToDoApp
Dockerfile

Git
repository
My Your
server server

Build Build

ToDoApp ToDoApp
Docker image Docker image
Figure 1.6 Building a
Docker application

Figure 1.6 shows what we’ll achieve by doing this.

The details of the application are unimportant. We’re going to demonstrate that
from the single short Dockerfile we’re about to give you, you can reliably build, run,
stop, and start an application in the same way on both your host and ours without
needing to worry about application installations or dependencies. This is a key part of
what Docker gives us—reliably reproduced and easily managed and shared develop-
ment environments. This means no more complex or ambiguous installation instruc-
tions to follow and potentially get lost in.

THE TO-DO APPLICATION This to-do application will be used a few times through-
out the book, and it’s quite a useful one to play with and demonstrate, so it’s
worth familiarizing yourself with it.

1.2.1 Ways to create a new Docker image

There are four standard ways to create Docker images. Table 1.2 itemizes these
methods.

Table 1.2 Options for creating Docker images

Method Description See technique

Docker commands / “By Fire up a container with docker run and input the See technique 14.
hand” commands to create your image on the command
line. Create a new image with docker commit.

Dockerfile Build from a known base image, and specify build Discussed shortly.
with a limited set of simple commands.

Licensed to Marocco Aurélien <[email protected]>

44 CHAPTER 1 Discovering Docker

Table 1.2 Options for creating Docker images (continued)

Method Description See technique

Dockerfile and configuration Same as Dockerfile, but hand over control of the See technique 47.
management (CM) tool build to a more sophisticated CM tool.

Scratch image and import a From an empty image, import a TAR file with the See technique 10.
set of files required files.

The first “by hand” option is fine if you’re doing proofs of concept to see whether
your installation process works. At the same time, you should be keeping notes about
the steps you’re taking so that you can return to the same point if you need to.
At some point you’re going to want to define the steps for creating your image.
This is the second option (and the one we’ll use here).
For more complex builds, you may want to go for the third option, particularly
when the Dockerfile features aren’t sophisticated enough for your image’s needs.
The final option builds from a null image by overlaying the set of files required to
run the image. This is useful if you want to import a set of self-contained files created
elsewhere, but it’s rarely seen in mainstream use.
We’ll look at the Dockerfile method now; the other methods will be covered later
in the book.

1.2.2 Writing a Dockerfile

A Dockerfile is a text file with a series of commands in it. Here’s the Dockerfile we’re
going to use for this example:

B Define the
base image. C Declare the
Clone the D FROM node maintainer.
todoapp MAINTAINER [email protected]
code. RUN git clone -q https://github.com/docker-in-practice/todo.git
WORKDIR todo
RUN npm install > /dev/null Run the node package
Move to the EXPOSE 8000 manager’s install
new cloned
directory. E
CMD ["npm","start"]
F
command (npm).

Specify which command

Specify that containers from the built
will be run on startup. H G image should listen on this port.

You begin the Dockerfile by defining the base image with the FROM command B. This
example uses a Node.js image so you have access to the Node.js binaries. The official
Node.js image is called node.
Next, you declare the maintainer with the MAINTAINER command C. In this case,
we’re using one of our email addresses, but you can replace this with your own

Licensed to Marocco Aurélien <[email protected]>

 Building a Docker application 45

reference because it’s your Dockerfile now. This line isn’t required to make a
working Docker image, but it’s good practice to include one. At this point, the build
has inherited the state of the node container, and you’re ready to work on top of it.
Next, you clone the todoapp code with a RUN command D. This uses the specified
command to retrieve the code for the application, running git within the container.
Git is installed inside the base node image in this case, but you can’t take this kind of
thing for granted.
Now you move to the new cloned directory with a WORKDIR command E. Not only
does this change directory within the build context, but the last WORKDIR command
determines which directory you’re in by default when you start up your container
from your built image.
Next, you run the node package manager’s install command (npm) F. This will set
up the dependencies for your application. You aren’t interested in the output here, so
you redirect it to /dev/null.
Because port 8000 is used by the application, you use the EXPOSE command to tell
Docker that containers from the built image should listen on this port G.
Finally, you use the CMD command to tell Docker which command will be run on
startup of the container H.
This simple example illustrates several key features of Docker and Dockerfiles. A
Dockerfile is a simple sequence of a limited set of com-
mands run in strict order. They affect the files and meta- The docker Path to the
command Dockerfile file
data of the resulting image. Here the RUN command
affects the filesystem by checking out and installing appli-
cations, and the EXPOSE, CMD, and WORKDIR commands
docker build .
affect the metadata of the image.

1.2.3 Building a Docker image

You’ve defined your Dockerfile’s build steps. Now The docker build
subcommand
you’re going to build the Docker image from it by typ-
ing the command in figure 1.7. Figure 1.7 Docker build
The output you’ll see will be similar to this: command

Docker uploads
the files and Each build step is
directories under numbered sequentially
the path supplied Sending build context to Docker daemon 178.7 kB
Sending build context to Docker daemon from 0 and output with
to the docker the command.
build command. Step 0 : FROM node
---> fc81e574af43
Step 1 : MAINTAINER [email protected]
---> Running in 21af1aad6950
To save space, each
Each command intermediate container is
---> 8f32669fe435
results in a new removed before continuing.
Removing intermediate container 21af1aad6950
image being Step 2 : RUN git clone https://github.com/ianmiell/todo.git
created, and the ---> Running in 0a030ee746ea
image ID is output. Cloning into 'todo'...

Licensed to Marocco Aurélien <[email protected]>

46 CHAPTER 1 Discovering Docker

---> 783c68b2e3fc
Removing intermediate container 0a030ee746ea
Step 3 : WORKDIR todo
---> Running in 2e59f5df7152
---> 8686b344b124
Removing intermediate container 2e59f5df7152
Step 4 : RUN npm install
Debug of the build is
output here (and edited
---> Running in bdf07a308fca
out of this listing).
npm info it worked if it ends with ok
[...]
npm info ok
---> 6cf8f3633306
Removing intermediate container bdf07a308fca
Step 5 : RUN chmod -R 777 /todo
---> Running in c03f27789768
---> 2c0ededd3a5e
Removing intermediate container c03f27789768
Step 6 : EXPOSE 8000
---> Running in 46685ea97b8f
---> f1c29feca036
Removing intermediate container 46685ea97b8f
Step 7 : CMD npm start
---> Running in 7b4c1a9ed6af
---> 439b172f994e
Final image ID for this
Removing intermediate container 7b4c1a9ed6af
build, ready to tag
Successfully built 439b172f994e

You now have a Docker image with an image ID

The docker The image
(“66c76cea05bb” in the preceding example, command ID to tag
but your ID will be different). It can be cumber-
some to keep referring to this ID, so you can tag
it for easier reference. docker tag 66c76cea05bb todoapp
Type the preceding command, replacing
the 66c76cea05bb with whatever image ID was
generated for you.
The docker tag Tag name to
You can now build your own copy of a Docker subcommand give image
image from a Dockerfile, reproducing an envi-
Figure 1.8 Docker tag command
ronment defined by someone else!

1.2.4 Running a Docker container

You’ve built and tagged your Docker image. Now you can run it as a container:

docker run -p 8000:8000 --name example1 todoapp The docker run

The output npm install subcommand starts the
of the npm info it worked if it ends with ok container, -p maps the
container’s npm info using [email protected] container’s port 8000 to
starting npm info using [email protected] the port 8000 on the host
process is npm info prestart [email protected] machine, --name gives the
sent to the container a unique name,
terminal. > [email protected] prestart /todo and the last argument is
> make all B the image.

Licensed to Marocco Aurélien <[email protected]>

 Building a Docker application 47

npm install
npm info it worked if it ends with ok
npm info using [email protected]
npm info using [email protected]
npm WARN package.json [email protected] No repository field.
npm WARN package.json [email protected] license should be a
➥ valid SPDX license expression
npm info preinstall [email protected]
npm info package.json [email protected] license should be a valid
➥ SPDX license expression
npm info package.json [email protected] No license field.
npm info package.json [email protected] No license field.
npm info package.json [email protected] license should be a valid
➥ SPDX license expression
npm info package.json [email protected] No license field.
npm info build /todo
npm info linkStuff [email protected]
npm info install [email protected]
npm info postinstall [email protected]
npm info prepublish [email protected]
npm info ok
if [ ! -e dist/ ]; then mkdir dist; fi
cp node_modules/react/dist/react.min.js dist/react.min.js

LocalTodoApp.js:9: // TODO: default english version

LocalTodoApp.js:84: fwdList =
➥ this.host.get('/TodoList#'+listId); // TODO fn+id sig
TodoApp.js:117: // TODO scroll into view
TodoApp.js:176: if (i>=list.length()) { i=list.length()-1; }
➥ // TODO .length
local.html:30: <!-- TODO 2-split, 3-split -->
model/TodoList.js:29:
➥ // TODO one op - repeated spec? long spec?
view/Footer.jsx:61: // TODO: show the entry's metadata
view/Footer.jsx:80: todoList.addObject(new TodoItem());
➥ // TODO create default
view/Header.jsx:25:
➥ // TODO list some meaningful header (apart from the id)
npm info start [email protected]
Hit Ctrl-C C
here to > [email protected] start /todo
terminate > node TodoAppServer.js Run this command to see
the process D containers that have been
and the Swarm server started port 8000 started and removed, along with
container. ^C an ID and status (like a process).
$ docker ps -a
CONTAINER ID IMAGE COMMAND CREATED
➥ STATUS PORTS NAMES
b9db5ada0461 todoapp:latest "npm start" 2 minutes ago
➥ Exited (130) 2 minutes ago example1
$ docker start example1 Restart the container,
this time in the
E background.

Licensed to Marocco Aurélien <[email protected]>

 48 CHAPTER 1 Discovering Docker

Run the ps F example1

command $ docker ps -a
again to CONTAINER ID IMAGE COMMAND CREATED
see the ➥ STATUS PORTS NAMES
changed b9db5ada0461 todoapp:latest "npm start" 8 minutes ago
status.
➥ Up 10 seconds 0.0.0.0:8000->8000/tcp example1
$ docker diff example1
The docker diff C /todo The /todo directory
subcommand A /todo/.swarm
H
has been changed.
shows you what A /todo/.swarm/TodoItem
files have been A /todo/.swarm/TodoItem/1tlOc02+A~4UZcz
affected since A /todo/.swarm/_log
The /todo/.swarm directory
the image was A /todo/dist
instantiated as A /todo/dist/LocalTodoApp.app.js
I
has been added.
a container. G A /todo/dist/TodoApp.app.js
A /todo/dist/react.min.js

The docker run subcommand starts up the container B. The -p flag maps the con-
tainer’s port 8000 to the port 8000 on the host machine, so you should now be able to
navigate with your browser to http://localhost:8000 to view the application. The - -name
flag gives the container a unique name you can refer to later for convenience. The last
argument is the image name.
Once the container was started, we hit CTRL-C to terminate the process and the
container C. You can run the ps command to see the containers that have been
started but not removed D. Note that each container has its own container ID and sta-
tus, analogous to a process. Its status is Exited, but you can restart it E. After you do,
notice how the status has changed to Up and the port mapping from container to host
machine is now displayed F.
The docker diff subcommand shows you which files have been affected since the
image was instantiated as a container G. In this case, the todo directory has been
changed H and the other listed files have been added I. No files have been deleted,
which is the other possibility.
As you can see, the fact that Docker “contains” your environment means that you
can treat it as an entity on which actions can be predictably performed. This gives
Docker its breadth of power—you can affect the software lifecycle from development
to production and maintenance. These changes are what this book will cover, showing
you in practical terms what can be done with Docker.
Next you’re going to learn about layering, another key concept in Docker.

1.2.5 Docker layering

Docker layering helps you manage a big problem that arises when you use containers
at scale. Imagine what would happen if you started up hundreds—or even thou-
sands—of the to-do app, and each of those required a copy of the files to be stored
somewhere.
As you can imagine, disk space would run out pretty quickly! By default, Docker
internally uses a copy-on-write mechanism to reduce the amount of disk space required

Licensed to Marocco Aurélien <[email protected]>

 Building a Docker application 49

Copy-on-startup Copy-on-write layers

ToDoApp ToDoApp ToDoApp ToDoApp

ToDoApp ToDoApp ToDoApp

ToDoApp ToDoApp ToDoApp

A non-layered application Each block represents a running

with nine copies made on disk container’s file differences from
for nine running instances. the original ToDoApp’s image. Figure 1.9 Copy-on-
Uses much less disk space. startup vs copy-on-write

(see figure 1.9). Whenever a running container needs to write to a file, it records the
change by copying the item to a new area of disk. When a Docker commit is performed,
this new area of disk is frozen and recorded as a layer with its own identifier.
This partly explains how Docker containers can start up so quickly—they have
nothing to copy because all the data has already been stored as the image.

COPY-ON-WRITE Copy-on-write is a standard optimization strategy used in com-

puting. When you create a new object (of any type) from a template, rather than
copying the entire set of data required, you only copy data over when it’s
changed. Depending on the use case, this can save considerable resources.

Figure 1.10 illustrates that the to-do app you’ve built has three layers you’re interested in.

Name: todoapp
Image ID: bd0921d1
Size: 600k Your to-do app’s
files on top
Name: node
Image ID: efc12dea
Three layers together Size: 1.5M Node binaries and files
make the to-do image. are added in this layer.

Name: Ubuntu The base Ubuntu

Image ID: abcf13de layer contains standard
Image IDs are unique Size: 89M base packages for the
references for images. distribution.

Figure 1.10 The to-do app’s filesystem layering in Docker

Licensed to Marocco Aurélien <[email protected]>

50 CHAPTER 1 Discovering Docker

Because the layers are static, you only need build on top of the image you wish to take
as a reference, should you need anything to change in a higher layer. In the to-do app,
you built from the publicly available node image and layered changes on top.
All three layers can be shared across multiple running containers, much as a
shared library can be shared in memory across multiple running processes. This is a
vital feature for operations, allowing the running of numerous containers based on
different images on host machines without running out of disk space.
Imagine that you’re running the to-do app as a live service for paying customers.
You can scale up your offering to a large number of users. If you’re developing, you
can spin up many different environments on your local machine at once. If you’re
moving through tests, you can run many more tests simultaneously, and far more
quickly than before. All these things are made possible by layering.
By building and running an application with Docker, you’ve begun to see the
power that Docker can bring to your workflow. Reproducing and sharing specific envi-
ronments and being able to land these in various places gives you both flexibility and
control over development.

1.3 Summary
Depending on your previous experience with Docker, this chapter might have been a
steep learning curve. We’ve covered a lot of ground in a short time.
You should now
Understand what a Docker image is
Know what Docker layering is, and why it’s useful
Be able to commit a new Docker image from a base image
Know what a Dockerfile is
We’ve used this knowledge to
Create a useful application
Reproduce state in an application with minimal effort
Next we’re going to introduce techniques that will help you understand how Docker
works and, from there, discuss some of the broader technical debate around Docker’s
usage. These first two introductory chapters form the basis for the remainder of the
book, which will take you from development to production, showing you how Docker
can be used to improve your workflow.

Licensed to Marocco Aurélien <[email protected]>

 Summary 51

Docker is impossible to ignore. This lightweight con-

tainer system is easier to deploy and more flexible than
traditional VMs. Built for simplicity and speed, it radi-
cally reduces your reliance on manual system adminis-
tration for tasks like configuring servers, creating
disposable (and portable!) development environments,
and predictably rolling out applications on unknown
systems. While the idea behind Docker is simple, it can
have a major impact on how you develop and deploy
software.
Docker in Practice is a hands-on guide to over 100
specific techniques you can use to get the most out of
Docker in your daily work. Following a cookbook-style Problem/Solution/Discussion
format, this practical handbook gives you instantly-useful solutions for important areas
like effortless server maintenance and configuration, deploying microservices, creat-
ing safe environments for experimentation, and much more. As you read, you'll grad-
uate from Docker basics into must-have practices like integrating Docker with your
Continuous Integration process, automating complex container creation with Chef,
and orchestration with Kubernetes.

What's inside
Speed up your DevOps pipeline with the use of containers
Reduce the effort of maintaining and configuring software
Using Docker to cheaply replace VMs
Streamlining your cloud workflow
Using the Docker Hub and its workflow
Navigating the Docker ecosystem
Written for developers and devops engineers who have already started their Docker
journey and want to use it effectively in a production setting.

Licensed to Marocco Aurélien <[email protected]>

 Mesos in Action

M anaging a growing cloud infrastructure can become a complex task

especially if you need to operate using multiple cloud providers as well as on-site
environments. Mesos adds a layer of abstraction on top of your infrastructure.
This means you are able to manage whole infrastructures as if everything is run-
ning on a single machine. Mesos in Action introduces you to the Apache Mesos
cluster manager and the concept of application-centric infrastructure. Chapter
1, “Introducing Mesos,” gives you a look behind the scenes of the abstraction
layer.

Licensed to Marocco Aurélien <[email protected]>

 Chapter 1 from Mesos in Action by Roger
Ignazio.

Introducing Mesos

This chapter covers

Introducing Mesos
Comparing Mesos with a traditional datacenter
Understanding when and why to use Mesos
Working with Mesos’s distributed architecture

Traditionally, physical—and virtual—machines have been the typical units of com-

puting in a datacenter. Machines are provisioned with various configuration man-
agement tools to later have applications deployed. These machines are usually
organized into clusters providing individual services, and systems administrators
oversee their day-to-day operations. Eventually, these clusters reach their maximum
capacity, and more machines are brought online to handle the load.
In 2010, a project at the University of California, Berkeley, aimed to solve the
scaling problem. The software project, now known as Apache Mesos, abstracts CPU,
memory, and disk resources in a way that allows datacenters to function as if they
were one large machine. Mesos creates a single underlying cluster to provide appli-

53

Licensed to Marocco Aurélien <[email protected]>

54 CHAPTER 1 Introducing Mesos

cations with the resources they need, without the overhead of virtual machines and
operating systems. You can see a simplified example of this in figure 1.1.

Mesos offers
available cluster
Spark Jenkins CI Marathon Chronos resources directly
to frameworks.

Mesos Mesos, like

the OS kernel,
abstracts resources.

OS OS OS OS OS OS OS
The OS kernel
provides access to
underlying physical
or virtual resources
Server Server Server Server Server Server Server
(CPU, memory, disk)

Figure 1.1 Frameworks sharing datacenter resources offered by Mesos

This book introduces Apache Mesos, an open source cluster manager that allows sys-
tems administrators and developers to focus less on individual servers and more on
the applications that run on them. You’ll see how to get up and running with Mesos in
your environment, how it shares resources and handles failure, and—perhaps most
important—how to use it as a platform to deploy applications.

1.1 Meet Mesos

Mesos works by introducing a layer of abstraction that provides a means to use entire
datacenters as if they were a single, large server. Instead of focusing on one applica-
tion running on a specific server, Mesos’s resource isolation allows for multitenancy—
the ability to run multiple applications on a single machine—leading to more effi-
cient use of computing resources.
To better understand this concept, you might think of Mesos as being similar to
today’s virtualization solutions: just as a hypervisor abstracts physical CPU, memory,
and storage resources and presents them to virtual machines, Mesos does the same
but offers these resources directly to applications. Another way to think about this is in
the context of multicore processors: when you launch an application on your laptop,
it runs on one or more cores, but in most cases it doesn’t particularly matter which
one. Mesos applies this same concept to datacenters.
In addition to improving overall resource use, Mesos is distributed, highly available,
and fault-tolerant right out of the box. It has built-in support for isolating processes using
containers, such as Linux control groups (cgroups) and Docker, allowing multiple appli-
cations to run alongside each other on a single machine. Where you once might have set
up three clusters—one each to run Memcached, Jenkins CI, and your Ruby on Rails
apps—you can instead deploy a single Mesos cluster to run all of these applications.

Licensed to Marocco Aurélien <[email protected]>

 Meet Mesos 55

In the next few sections, you’re going to look at how Mesos works to provide all of
these features and how it compares to a traditional datacenter.

1.1.1 Understanding how it works

Using a combination of concepts referred to as resource offers, two-tier scheduling,
and resource isolation, Mesos provides a means for the cluster to act as a single super-
computer on which to run tasks. Before digging in too deeply here, let’s take a look at
figure 1.2. This diagram demonstrates the logic Mesos follows when offering resources
to running applications. This particular example references the Apache Spark data-
processing framework.

Mesos slave

Mesos master
Container

Spark scheduler:
User-submitted Spark job
Do I have work to do?

Yes No

Accept Reject

Figure 1.2 Mesos advertises the available CPU, memory, and disk as resource offers to
frameworks.

Let’s break it down:

B The Mesos slave offers its available CPU, memory, and disk to the Mesos master
in the form of a resource offer.
c The Mesos master’s allocation module—or scheduling algorithm—decides which
frameworks—or applications—to offer the resources to.
d In this particular case, the Spark scheduler doesn’t have any jobs to run on the
cluster. It rejects the resource offer, allowing the master to offer the resources to
another framework that might have some work to do.
e Now consider a user submitting a Spark job to be run on the cluster. The sched-
uler accepts the job and waits for a resource offer that satisfies the workload.
f The Spark scheduler accepts a resource offer from the Mesos master, and
launches one or more tasks on an available Mesos slave. These tasks are launched

Licensed to Marocco Aurélien <[email protected]>

56 CHAPTER 1 Introducing Mesos

within a container, providing isolation between the various tasks that might be
running on a given Mesos slave.

Seems simple, right? Now that you’ve learned how Mesos uses resource offers to adver-
tise resources to frameworks, and how two-tier scheduling allows frameworks to accept
and reject resource offers as needed, let’s take a closer look at some of these funda-
mental concepts.

NOTE An effort is underway to rename the Mesos slave role to agent for future
versions of Mesos. Because this book covers Mesos 0.22.2, it uses the terminol-
ogy of that specific release, so as to not create any unnecessary confusion. For
more information, see https://issues.apache.org/jira/browse/MESOS-1478.

RESOURCE OFFERS
Like many other cluster managers, Mesos clusters are made up of groups of machines
called masters and slaves. Each Mesos slave in a cluster advertises its available CPU,
memory, and storage in the form of resource offers. As you saw in figure 1.2, these
resource offers are periodically sent from the slaves to the Mesos masters, processed
by a scheduling algorithm, and then offered to a framework’s scheduler running on
the Mesos cluster.
TWO-TIER SCHEDULING
In a Mesos cluster, resource scheduling is the responsibility of the Mesos master’s allo-
cation module and the framework’s scheduler, a concept known as two-tier scheduling.
As previously demonstrated, resource offers from Mesos slaves are sent to the master’s
allocation module, which is then responsible for offering resources to various frame-
work schedulers. The framework schedulers can accept or reject the resources based
on their workload.
The allocation module is a pluggable component of the Mesos master that
implements an algorithm to determine which offers are sent to which frameworks
(and when). The modular nature of this component allows systems engineers to
implement their own resource-sharing policies for their organization. By default,
Mesos uses an algorithm developed at UC Berkeley known as Dominant Resource
Fairness (DRF):
In a nutshell, DRF seeks to maximize the minimum dominant share across all users. For
example, if user A runs CPU-heavy tasks and user B runs memory-heavy tasks, DRF
attempts to equalize user A’s share of CPUs with user B’s share of memory. In the single-
resource case, DRF reduces to max-min fairness for that resource.1

1
A. Ghodsi, M. Zaharia, B. Hindman, A. Konwinski, S. Shenker, and I. Stoica. “Dominant Resource Fairness:
Fair Allocation of Multiple Resource Types.” NSDI, vol. 11, 2011.

Licensed to Marocco Aurélien <[email protected]>

 Meet Mesos 57

Mesos’s use of the DRF algorithm by default is fine for most deployments. Chances are
you won’t need to write your own allocation algorithm, so this book doesn’t go into
much detail about DRF. If you’re interested in learning more about this research, you
can find the paper online at www.usenix.org/legacy/events/nsdi11/tech/full_pa-
pers/Ghodsi.pdf.
RESOURCE ISOLATION
Using Linux cgroups or Docker containers to isolate processes, Mesos allows for mult-
itenancy, or for multiple processes to be executed on a single Mesos slave. A framework
then executes its tasks within the container, using a Mesos containerizer. If you’re not
familiar with containers, think of them as a lightweight approach to how a hypervisor
runs multiple virtual machines on a single physical host, but without the overhead or
need to run an entire operating system.

NOTE In addition to Docker and cgroups, Mesos provides another means of

isolation for other POSIX-compliant operating systems: posix/cpu,
posix/mem, and posix/disk. It’s worth noting that these isolation methods
don’t isolate resources, but instead monitor resource use.

Now that you have a clearer understanding of how Mesos works, you can move on to
understanding how this technology compares to the traditional datacenter. More spe-
cifically, the next section introduces the concept of an application-centric datacenter,
where the focus is more on applications than on the servers and operating systems
that run them.

1.1.2 Comparing virtual machines and containers

When thinking about applications deployed in a traditional datacenter, virtual
machines often come to mind. In recent years, virtualization providers (VMware,
OpenStack, Xen, and KVM, to name a few) have become commonplace in many orga-
nizations. Similar to how a hypervisor allows a physical host’s resources to be
abstracted and shared among virtual machines, Mesos provides a layer of abstraction,
albeit at a different level. The resources are presented to applications themselves, and
in turn consumed by containers.
To illustrate this point, consider figure 1.3, which compares the various layers of
infrastructure required to deploy four applications.
VIRTUAL MACHINES
When thinking about traditional virtual machine–based application deployments,
consider for a moment the operational overhead of maintaining the operating sys-
tems on each of them: installing packages, applying security updates, maintaining
user access, identifying and remediating configuration drift; the list goes on. What’s
the added benefit of running applications atop an entire operating system when
you’re more concerned with deploying the application itself? Not to mention the

Licensed to Marocco Aurélien <[email protected]>

58 CHAPTER 1 Introducing Mesos

Virtual machine–based
application deployment
Container-based
App App App App application deployment

OS OS OS OS App App App App

VM VM VM VM Mesos

Hypervisor Hypervisor OS OS

Physical server Physical server Physical server Physical server

Isolates apps by running multiple Isolates apps using

VMs per physical server; still need features of the host OS,
to manage each guest OS! such as Linux cgroups.

Figure 1.3 Comparing VM-based and container-based application deployments

overhead of the operating system, which consumes added CPU, memory, and disk. At
a large-enough scale, this becomes wasteful. With an application-centric approach to
managing datacenters, Mesos allows you to simplify your stack—and your application
deployments—using lightweight containers.
CONTAINERS
As you learned previously, Mesos uses containers for resource isolation between
processes. In the context of Mesos, the two most important resource-isolation meth-
ods to know about are the control groups (cgroups) built into the Linux kernel,
and Docker.
Around 2007, support for control groups (referred to as cgroups throughout this
text) was made available in the Linux kernel, beginning with version 2.6.24. This
allows the execution of processes in a way that’s sandboxed from other processes. In the
context of Mesos, cgroups provide resource constraints for running processes, ensur-
ing that they don’t interfere with other processes running on the system. When using
cgroups, any packages or libraries that the tasks might depend on (a specific version
of Python, a working C++ compiler, and so on) must be already present on the host
operating system. If your workloads, packages, and required tools and libraries are
fairly standardized or don’t readily conflict with each other, this might not be a prob-
lem. But consider figure 1.4, which demonstrates how using Docker can overcome
these sorts of problems and allow you to run applications and workloads in a more iso-
lated manner.

Licensed to Marocco Aurélien <[email protected]>

 Meet Mesos 59

Virtual machines Docker containers Linux control groups

Application

Application
Packages and
libraries
Application
Packages and
libraries
Guest OS Linux cgroup
Docker container

Packages and
VM Docker engine libraries

Hypervisor Host OS Host OS

Figure 1.4 Comparing virtual machines, Docker containers, and

Linux cgroups

Using low-level primitives in the Linux kernel, including cgroups and namespaces,
Docker provides a means to build and deploy containers almost as if they were virtual
machines. The application and all of its dependencies are packaged within the con-
tainer and deployed atop a host operating system. They take a concept from the
freight industry—the standardized industrial shipping container—and apply this to
application deployment. In recent years, this new unit of software delivery has grown
in popularity as it’s generally considered to be more lightweight than deploying an
entire virtual machine.
You don’t need to understand all the implementation details and intricacies of
building and deploying containers to use Mesos, though. If you’d like more informa-
tion, please consult the following online resources:
Linux control groups: www.kernel.org/doc/documentation/cgroup-v1/cgrou-
ps.txt
Docker: https://docs.docker.com

1.1.3 Knowing when (and why) to use Mesos

Running applications at scale isn’t reserved for large enterprises anymore. Startups
with only a handful of employees are creating apps that easily attract millions of users.
Re-architecting applications and datacenters is a nontrivial task, but certain compo-
nents that are in a typical stack are already great candidates to run on Mesos. By tak-
ing some of these technologies and moving them (and their workloads) to a Mesos
cluster, you can scale them more easily and run your datacenter more efficiently.

NOTE This book covers Mesos version 0.22.2, which provides an environment
for running stateless and distributed applications. Beginning in version 0.23,

Licensed to Marocco Aurélien <[email protected]>

60 CHAPTER 1 Introducing Mesos

Mesos will begin work to support persistent resources, thus enabling support
for stateful frameworks. For more information on this effort, see
https://issues.apache.org/jira/browse/MESOS-1554.

For example, consider the stateless, distributed, and stateful technologies in table 1.1.
Table 1.1 Technologies that are—and aren’t—good candidates to run on Mesos

Service type Examples Should you use Mesos?

Stateless—no need to persist Web apps (Ruby on Rails, Play, Yes

data to disk Django), Memcached, Jenkins CI
build slaves

Distributed out of the box Cassandra, Elasticsearch, Hadoop Yes, provided the correct level of
Distributed File System (HDFS) redundancy is in place

Stateful—needs to persist data MySQL, PostgreSQL, Jenkins CI No (version 0.22); potentially

to disk masters (version 0.23+)

The real value of Mesos is realized when running stateless services and applications—
applications that will handle incoming loads but that could go offline at any time with-
out negatively impacting the service as a whole, or services that run a job and report
the result to another system. As noted previously, examples of some of these applica-
tions include Ruby on Rails and Jenkins CI build slaves.
Progress has been made running distributed databases (such as Cassandra and
Elasticsearch) and distributed filesystems (such as Hadoop Distributed File System, or
HDFS) as Mesos frameworks. But this is feasible only if the correct level of redundancy
is in place. Although certain distributed databases and filesystems have data replica-
tion and fault tolerance built in, your data might not survive if the entire Mesos cluster
fails (because of natural disasters, redundant power/cooling systems failures, or human
error). In the real world, you should weigh the risks and benefits of deploying services
that persist data on a Mesos cluster.
As I mentioned earlier, Mesos excels at running stateless, distributed services.
Stateful applications that need to persist data to disk aren’t good candidates for run-
ning on Mesos as of this writing. Although possible, it’s not yet advisable to run certain
databases such as MySQL and PostgreSQL atop a Mesos cluster. When you do need to
persist data, it’s preferable to do so by deploying a traditional database cluster outside
the Mesos cluster.

1.2 Why we need to rethink the datacenter

Deploying applications within a datacenter has traditionally involved one or more
physical (or virtual) servers. The introduction and mainstream adoption of virtualiza-
tion has allowed us to run multiple virtual machines on a single physical server and
make better use of physical resources. But running applications this way also means
you’re usually running a full operating system on each of those virtual machines,
which consumes resources and brings along its own maintenance overhead.

Licensed to Marocco Aurélien <[email protected]>

 Why we need to rethink the datacenter 61

This section presents two primary reasons that you should rethink how datacenters
are managed: the administrative overhead of statically partitioning resources, and the
need to focus more on applications instead of infrastructure.

1.2.1 Partitioning of resources

When you consider the traditional virtual machine–based model of deploying applica-
tions and statically partitioning clusters, you quickly find this deployment model
inefficient and cumbersome to maintain. By maximizing the use of each server in a
datacenter, operations teams maximize their return on investment and can keep the
total cost of ownership as reasonable as possible.
In computing circles, teams generally refer to a cluster as a group of servers that
work together as a single system to provide a service. Traditionally, the deployment of
these services has been largely node-centric: you dedicate a certain number of machines
to provide a given service. But as the infrastructure footprint expands and service
offerings increase, it’s difficult to continue statically partitioning these services.
But now consider the demand for these services doubling. To continue scaling, a
systems administrator needs to provision new machines and join them to the individ-
ual clusters. Perhaps the operations team, anticipating the need for additional capac-
ity, scales each of those clusters to three times its current size. Although you’ve
managed to scale each of those services, you now have machines in your datacenter
sitting idle, waiting to be used. As such, if a single machine in any of those clusters
fails, it quickly needs to be brought back online for the service to continue operating
at full capacity, as shown in figure 1.5.

Spark slave Spark slave Spark slave

Each of these services

has been statically
Jenkins slave Jenkins slave Jenkins slave
partitioned onto
multiple servers.

X X Ruby on Rails
app server

If two of the app servers go down,

the Rails app can only operate at
one-third capacity, even if Spark
and Jenkins sit idle.

Figure 1.5 Three applications statically partitioned in a datacenter

Licensed to Marocco Aurélien <[email protected]>

62 CHAPTER 1 Introducing Mesos

Now consider solving the aforementioned scaling scenario by using Mesos, as shown
in figure 1.6. You can see that you’d use these same machines in the datacenter to
focus on running applications instead of virtual machines. The applications could run
on any machine with available resources. If you need to scale, you add servers to the
Mesos cluster, instead of adding machines to multiple clusters. If a single Mesos node
goes offline, no particular impact occurs to any one service.

These services are run on Mesos,

which dynamically schedules them
within the cluster based on
available capacity.

X X Mesos slave

Mesos slave Mesos slave Mesos slave

Mesos slave Mesos slave Mesos slave

Figure 1.6 Three applications running on a Mesos cluster

Consider these small differences across hundreds or thousands of servers. Instead of

trying to guess how many servers you need for each service and provision them into
several static clusters, you’re able to allow these services to dynamically request the
compute, memory, and storage resources they need to run. To continue scaling, you
add new machines to your Mesos cluster, and the applications running on the cluster
scale to the new infrastructure. Operating a single, large computing cluster in this
manner has several advantages:
You can easily provision additional cluster capacity.
You can be less concerned about where services are running.
You can scale from several nodes to thousands.
The loss of several servers doesn’t severely degrade any one service.

1.2.2 Deploying applications

As we discussed previously, one of the major differences—and benefits—of deploying
applications on a Mesos cluster is multitenancy. Not unlike a virtualization hypervisor
running multiple virtual machines on a physical server, Mesos allows multiple applica-
tions to run on a single server in isolated environments, using either Linux cgroups or

Licensed to Marocco Aurélien <[email protected]>

 The Mesos distributed architecture 63

Docker containers. Instead of having multiple environments (one each for develop-
ment, staging, and production), the entire datacenter becomes a platform on which
to deploy applications.
Where Mesos is commonly referred to—and acts as—a distributed kernel, other
Mesos frameworks help users run long-running and scheduled tasks, similar to the
init and Cron systems, respectively. You’ll learn more about these frameworks (Mar-
athon, Chronos, and Aurora) and how to deploy applications on them later in
this book.
Consider the power of what I’ve described so far: Mesos provides fault tolerance
out of the box. Instead of a systems administrator getting paged when a single server
goes offline, the cluster will automatically start the failed job elsewhere. The sysadmin
needs to be concerned only if a certain percentage of machines goes offline in the
datacenter, as that might signal a larger problem. As such, with the correct placement
and redundancy in place, scheduled maintenance can occur at any time.

1.3 The Mesos distributed architecture

To provide services at scale, Mesos provides a distributed, fault-tolerant architecture
that enables fine-grained resource scheduling. This architecture comprises three com-
ponents: masters, slaves, and the applications (commonly referred to as frameworks) that
run on them. Mesos relies on Apache ZooKeeper, a distributed database used specifi-
cally for coordinating leader election within the cluster, and for leader detection by
other Mesos masters, slaves, and frameworks.
In figure 1.7, you can see how each of these architecture components works
together to provide a stable platform on which to deploy applications. I’ll break it
down for you in the sections that follow the diagram.

1.3.1 Masters
One or more Mesos masters are responsible for managing the Mesos slave daemons
running on each machine in the cluster. Using ZooKeeper, they coordinate which
node will be the leading master, and which masters will be on standby, ready to take
over if the leading master goes offline.
The leading master is responsible for deciding which resources to offer to a partic-
ular framework using a pluggable allocation module, or scheduling algorithm, to distrib-
ute resource offers to the various schedulers. The scheduler can then either accept or
reject the offer based on whether it has any work to be performed at that time.
A Mesos cluster requires a minimum of one master, and three or more are recom-
mended for production deployments to ensure that the services are highly available.
You can run ZooKeeper on the same machines as the Mesos masters themselves, or
use a standalone ZooKeeper cluster. Chapter 3 goes into more detail about the sizing
and deploying of Mesos masters.

Licensed to Marocco Aurélien <[email protected]>

64 CHAPTER 1 Introducing Mesos

Framework A Framework B
scheduler scheduler

Resource offer Resource offer

Mesos master quorum

Mesos master
(leader)

ZK

ZK ZK

Mesos master Mesos master

(standby) (standby)

Resource offer Resource offer

Slave 1 Slave n

Framework A Framework A Framework B

executor executor executor

Tasks Tasks Tasks Figure 1.7 The Mesos architecture

consists of one or more masters,
slaves, and frameworks.

1.3.2 Slaves
The machines in a cluster responsible for executing a framework’s tasks are referred
to as Mesos slaves. They query ZooKeeper to determine the leading Mesos master and
advertise their available CPU, memory, and storage resources to the leading master in
the form of a resource offer. When a scheduler accepts a resource offer from the
Mesos master, it then launches one or more executors on the slave, which are responsi-
ble for running the framework’s tasks.
Mesos slaves can also be configured with certain attributes and resources, which
allow them to be customized for a given environment. Attributes refer to key/value
pairs that might contain information about the node’s location in a datacenter, and
resources allow a particular slave’s advertised CPU, memory, and disk to be overridden
with user-provided values, instead of Mesos automatically detecting the available
resources on the slave. Consider the following example attributes and resources:
--attributes='datacenter:pdx1;rack:1-1;os:rhel7'
--resources='cpu:24;mem:24576;disk:409600'

Licensed to Marocco Aurélien <[email protected]>

 Summary 65

I’ve configured this particular Mesos slave to advertise its datacenter; location within
the datacenter; operating system; and user-provided CPU, memory, and disk resources.
This information is especially useful when trying to ensure that applications stay
online during scheduled maintenance. Using this information, a datacenter operator
could take an entire rack (or an entire row!) of machines offline for scheduled main-
tenance without impacting users. Chapter 4 covers this (and more) in the Mesos slave
configuration section.

1.3.3 Frameworks
As you learned earlier, a framework is the term given to any Mesos application that’s
responsible for scheduling and executing tasks on a cluster. A framework is made up
of two components: a scheduler and an executor.

TIP A list of frameworks known to exist at the time of writing is included in

appendix B.

SCHEDULER
A scheduler is typically a long-running service responsible for connecting to a Mesos
master and accepting or rejecting resource offers. Mesos delegates the responsibility
of scheduling to the framework, instead of attempting to schedule all the work for a
cluster itself. The scheduler can then accept or reject a resource offer based on
whether it has any tasks to run at the time of the offer. The scheduler detects the lead-
ing master by communicating with the ZooKeeper cluster, and then registers itself to
that master accordingly.
EXECUTOR
An executor is a process launched on a Mesos slave that runs a framework’s tasks on a
slave. As of this writing, the built-in Mesos executors allow frameworks to execute shell
scripts or run Docker containers. New executors can be written using Mesos’s various
language bindings and bundled with the framework, to be fetched by the Mesos slave
when a task requires it.
As you’ve learned, Mesos provides a distributed, highly available architecture. Mas-
ters schedule work to be performed on the cluster, and slaves advertise available
resources to the schedulers, which in turn execute tasks on the cluster.

1.4 Summary
In this chapter, you’ve been introduced to the Apache Mesos project, its architecture,
and how it attempts to solve scaling problems and make clustering simple. You’ve also
learned how Mesos deployments compare and contrast with the traditional datacen-
ter, and how an application-centric approach can lead to using resources more effi-
ciently. We’ve discussed when (and when not) to use Mesos for a given workload, and

Licensed to Marocco Aurélien <[email protected]>

66 CHAPTER 1 Introducing Mesos

where you can get help and find more information, should you need it. Here are a few
things to remember:
Mesos abstracts CPU, memory, and disk resources away from underlying systems
and presents multiple machines as a single entity.
Mesos slaves advertise their available CPUs, memory, and disk in the form of
resource offers.
A Mesos framework comprises two primary components: a scheduler and an
executor.
Containers are a lightweight method to provide resource isolation to individual
processes.
In the next chapter, I’ll walk you through a real-world example of how Mesos allows
for more efficient resource use, and how you might run applications in your own data-
center by building on projects in the Mesos ecosystem.

Licensed to Marocco Aurélien <[email protected]>

 Summary 67

The modern "data center" is a complex arena, with

physical and virtual servers, multiple OS environments,
and complex networking that frequently spans multi-
ple locations. When you throw Docker and container-
based systems like CoreOS and Project Atomic into the
mix, along with the increasingly important require-
ment for infrastructure automation, the need to sim-
plify the data center has never been greater. Mesos, an
innovative open-source cluster management platform,
transforms the whole data center into a single pool of
compute, memory and storage resources that you can
allocate, automate, and scale as if you're working with a
single super-computer. Mesos is an ideal environment for deploying containerized
applications at scale, and it's generating huge buzz in the big data world as a saner
environment for running Spark and Hadoop.
Mesos in Action introduces readers to the Apache Mesos cluster manager and the
concept of application-centric infrastructure. It guides you from your first steps in
deploying a highly available Mesos cluster through deploying applications in produc-
tion and writing native Mesos frameworks. You'll learn how to scale to thousands of
nodes, while providing resource isolation between processes using Linux and Docker
containers. You'll also learn practical techniques for deploying applications using pop-
ular key frameworks, including Marathon, Chronos, and Aurora. Along the way, you'll
get a good look into Mesos internals, including fault tolerance, slave attributes, and
resource scheduling and Mesos administration, including logging, monitoring, frame-
work authorization, and slave recovery.

What's inside
Spinning up your first Mesos cluster
Deploying containerized applications on Mesos
Scheduling, resource administration, and logging
Deploy applications using the popular Marathon, Chronos, and Aurora frame-
works
Writing custom Mesos frameworks using Python
Readers need to be familiar with the core ideas of data center administration, includ-
ing networking, virtualization, and application deployment on Linux systems. The
Python-based code examples should be clear to readers using Mesos bindings for
other popular languages, including C++, Go, and Scala.

Licensed to Marocco Aurélien <[email protected]>

 RabbitMQ in Depth

D ecoupling is a popular pattern when architecting systems for the cloud.

By decoupling different parts of your system, you end up with scalability and
high availability. RabbitMQ, a message-oriented middleware supports you when
creating distributed software architectures. RabbitMQ in Depth gives you valuable
insights into decoupling in general and RabbitMQ in particular. The first chap-
ter, “Foundational RabbitMQ,” covers features and benefits from using Rab-
bitMQ for your future cloud application.

Licensed to Marocco Aurélien <[email protected]>

 Chapter 1 from RabbitMQ in Depth by Gavin M.
Roy.

Foundational RabbitMQ

This chapter covers:

Unique features of RabbitMQ
Why RabbitMQ is becoming a popular choice for the
centerpiece of messaging-based architectures
The basics of the Advanced Messaging Queuing
Model, RabbitMQ’s foundation

Whether your application is in the cloud or in your own data center, RabbitMQ is a
lightweight and extremely powerful tool for creating distributed software architec-
tures that ranges from the very simple to the incredibly complex. In this chapter
you will learn how RabbitMQ, as messaging-oriented middleware, allows tremen-
dous flexibility in how you approach and solve problems. You will learn about how
some companies are using it and key features that make RabbitMQ one of the most
popular message brokers today.

69

Licensed to Marocco Aurélien <[email protected]>

70 CHAPTER 1 Foundational RabbitMQ

1.1 RabbitMQ’s features and benefits

RabbitMQ has many features and benefits, the most important ones outlined below:
Originally developed in a partnership between LShift, LTD, and Cohesive FT as
RabbitMQ Technologies, RabbitMQ is now owned by Pivotal Software Inc. and
released under the Mozilla Public License. As an open-source project written in
Erlang, RabbitMQ enjoys freedom and flexibility while leveraging the strength
of Pivotals support of the product. Developers and engineers in the RabbitMQ
community contribute enhancements and add-ons while Pivotal offers commer-
cial support and a stable home for ongoing product maturation.
As a message broker that implements platform and vendor neutral AMQP specifi-
cation, there are clients available for almost any programming language and on
all major computer platforms.
It is lightweight, requiring under 40 MB of RAM to run the core RabbitMQ appli-
cation, along with plugins such as the Management UI. Note that adding mes-
sages to queues can and will increase its memory usage.
Client libraries target most modern programming languages on multiple platforms,
and RabbitMQ is a compelling broker. There are no vendor or language lock-
ins when choosing how to write programs that talk to RabbitMQ. In fact, it is
not uncommon to see RabbitMQ used as the centerpiece between applications
written in different languages. RabbitMQ provides a useful bridge that allows
for languages like Java, Ruby, Python, PHP, JavaScript and C# to share data
across operating systems and environments.
RabbitMQ provides flexibility in controlling the trade-offs of reliable messaging with
message throughput and performance. Because it’s not a “one size fits all” type
of application, messages can designate if they should be persisted to disk prior
to delivery. If setup in a cluster, queues can be highly-available, spanning multi-
ple servers to ensure that messages are not lost in case of server failure.
Because not all network topologies and architectures are the same, RabbitMQ
provides for messaging in low-latency environments and plugins for higher-latency
environments such as the Internet. This allows for RabbitMQ to be clustered on the
same local network and share federated messages across multiple data-centers.
As a center point for application integrations, RabbitMQ provides a flexible
plugin system. As an example, there are third-party plugins for storing messages
directly into databases, using RabbitMQ directly for database writes.
In RabbitMQ, security is provided in multiple layers. Client connections can be
secured by enforcing SSL-only communication and client certificate validation.
User access can be managed at the virtual-host level, providing isolation of mes-
sages and resources at a high-level. In addition, access to configuration capabil-
ities, the reading from queues and writing to exchanges, is managed by regular
expression (regex) pattern matching. Finally, plugins can be used for integra-
tion into external authentication systems like LDAP.

Licensed to Marocco Aurélien <[email protected]>

 RabbitMQ’s features and benefits 71

While we will explore the features on this list in later chapters, I would like to focus on
the two most foundational features of RabbitMQ: the language it is programmed in
(Erlang), and the model it is based on (the Advanced Message Queuing Model), a
specification that defines much of the RabbitMQ lexicon and behavior.

1.1.1 RabbitMQ and Erlang

As a highly performant, stable, and message broker that clusters well, it is no surprise
that RabbitMQ has found a home in such mission-critical environments as the center-
piece of large scale messaging architectures. It was written in Erlang, the telco-grade,
functional programming language designed at the Ericsson Computer Science Labo-
ratory in the mid-to-late 1980’s. Erlang was designed to be a distributed, fault-tolerant,
soft real-time system for applications that require 99.999% uptime. As a language and
runtime system, Erlang focuses on lightweight processes that pass messages amongst
each other providing a high level of concurrency with no shared state.

REAL-TIME SYSTEM A real-time system is a hardware platform, software plat-

form, or a combination of both that has requirements defined for when it
must return a response from an event. A soft real-time system will sacrifice less
important deadlines for executing tasks in favor of more important ones.

Erlang’s design around concurrent processing and message passing makes it a natural
choice for a message broker like RabbitMQ: As an application, a message broker
maintains concurrent connections, routes messages, and manages their state. In addi-
tion, Erlang’s distributed communication architecture makes it a natural for Rab-
bitMQ’s clustering mechanism. Servers in a RabbitMQ cluster make use of Erlang’s
inter-process communication (IPC) system, offloading the functionality that many com-
peting message brokers have to implement to add clustering capabilities (figure 1.1).
Despite the advantages RabbitMQ has using Erlang, the Erlang environment can
be a stumbling block. If this is your first foray into Erlang, check out Appendix B: Just
Enough Erlang. In the appendix, you will learn enough to be confident in managing
RabbitMQ’s configuration files and you will learn how to use Erlang to gather infor-
mation about RabbitMQ’s current runtime state.

1.1.2 RabbitMQ and AMQP

When RabbitMQ was originally released in 2007, interoperability, performance, and
stability were the primary goals in mind during development, and RabbitMQ was one
of the first message brokers to implement the Advanced Message Queuing Protocol
(AMQP) specification. By all appearances, it set out to be the reference implementa-
tion. Split into two parts, the AMQP specification defines not only the wire protocol
for talking to RabbitMQ, but the logical model that outlines RabbitMQ’s core func-
tionality.

Licensed to Marocco Aurélien <[email protected]>

72 CHAPTER 1 Foundational RabbitMQ

Figure 1.1 RabbitMQ clusters use the native Erlang inter-process communication mechanism in
the Virtual Memory (VM) for cross-node communication, sharing state information and allowing for
messages to be published and consumed across the entire cluster.

NOTE There are multiple versions of the AMQP specification. For the pur-
poses this book, we will focus only on AMQP 0-9-1. While newer versions of
RabbitMQ support AMQP 1.0 as a plugin extension, the core RabbitMQ
architecture is more closely related to AMQP 0-8 and 0-9-1. The AMQP speci-
fication is primarily comprised of two documents, a top-level document that
describes both the AMQ model and the AMQ protocol and a more detailed
document that provides varying levels of information about every class,
method, property and field. More information about AMQP, including the
specification documents may be found at http://www.amqp.org.

There are multiple popular message brokers and messaging protocols, and it is impor-
tant that you consider the impact the protocol and broker will have on your applica-
tion. While RabbitMQ supports AMQP, it also supports other protocols, such as
MQTT, Stomp, and XMPP. RabbitMQ’s protocol neutrality and plugin extensibility
make it a good choice for multi-protocol application architectures when compared to
other popular message brokers.
It is RabbitMQ’s roots in the AMQP specification that outline its primary architec-
ture and communication methodologies. This is an important distinction when evalu-
ating RabbitMQ against other message brokers. As with AMQP, RabbitMQ set out to
be a vendor-neutral, platform-independent solution for the complex needs that mes-
saging oriented architectures demand, such as flexible message routing, configurable
message durability, and inter-datacenter communication, to name a few.

Licensed to Marocco Aurélien <[email protected]>

 Who is using RabbitMQ, and how? 73

1.2 Who is using RabbitMQ, and how?

As an open-source software package, RabbitMQ is rapidly gaining mainstream adop-
tion and powers some of the largest, most trafficked websites on the Internet. Today,
RabbitMQ is known to run in many different environments and at many different
types of companies and organizations:
Reddit, the popular online community, utilizes RabbitMQ heavily in the core of
their application platform, which serves billions of web pages per month. When
a user registers on the site, submits a news post, or votes on a link, a message is
published into RabbitMQ for asynchronous processing by consumer applica-
tions.
NASA chose RabbitMQ to be message-broker for Nebula, a centralized server
management platform for their server infrastructure that grew into the Open-
Stack platform, a very popular software platform for building private and public
cloud services.
RabbitMQ sits at the core of Agoura Games’ community-oriented online gam-
ing platform and routes large volumes of real-time single and multiplayer game
data and events.
For the Ocean Observations Initiative, RabbitMQ routes mission-critical physi-
cal, chemical, geological, and biological data to a distributed network of
research computers. The data, collected from sensors in the Southern, Pacific,
and Atlantic oceans, are integral to a National Science Foundation project that
involves building a large-scale network of sensors in the ocean and seafloor.
Rapportive, a GMail add-on that places detailed contact information inside the
inbox, uses RabbitMQ as the glue for its data processing systems. Billions of
messages pass through RabbitMQ monthly to provide data to Rapportive’s web-
crawling engine and analytics system, and to offload long-running operations
from its web servers.
MercadoLibre, the largest e-commerce ecosystem in Latin America, use Rab-
bitMQ at the heart of their Enterprise Service Bus (ESB) architecture, decou-
pling their data from tightly coupled applications. This allows for flexible
integrations with various components in their application architecture.
Google’s AdMob mobile advertising network used RabbitMQ at the core of their
RockSteady project to do real-time metrics analysis and fault-detection by fun-
neling a fire hose of messages through RabbitMQ into Esper, the complex-
event-processing system.
India’s biometric database system, Aandhaar, leverages RabbitMQ to process
data at various stages in its workflow, delivering data to their monitoring tools,
data warehouse, and their Hadoop based data processing system. Aandhaar is a
system for providing an online portable identity system for every single resident
of India, covering 1.2 billion people.

Licensed to Marocco Aurélien <[email protected]>

74 CHAPTER 1 Foundational RabbitMQ

As you can see, RabbitMQ is not only used by some of the largest sites on the Internet,
it has found its way into academia for large scale scientific research, and NASA found
it fitting to use RabbitMQ at the core of their network infrastructure management
stack. As these examples show, RabbitMQ has been used in mission-critical applica-
tions in many different environments and industries with tremendous success.

1.3 The advantages of loosely coupled architectures

When I first started to implement a messaging based architecture, I was looking for a
way to decouple needed database updates when a member logged into a website. The
website had grown very quickly and was not initially designed to scale well. As a user
logged into the website, several database servers needed to be updated with a time-
stamp when the member logged in (figure 1.2). This timestamp needed to be
updated in real-time, as the most engaging activities on the site were driven in part by
the timestamp value. Upon login, members were given preferential status in social
games to those users who were actively online at any given time.

Figure 1.2 Before: Once a user has logged in, each database is updated with a timestamp
sequentially and dependently. The more tables you add increases the time this takes.

Licensed to Marocco Aurélien <[email protected]>

 The advantages of loosely coupled architectures 75

As the site continued to grow, the amount of time it took for a member to login
also grew. The reason for this was fairly straightforward; when adding a new applica-
tion that used the member’s last login timestamp, its database tables would carry the
value to make it as fast as possible by removing cross database joins. To keep the data
up to date and accurate, the new data tables would also be updated when the member
logged in. It was not long before there were many tables being maintained this way.
The performance degradation began to creep up as the database updates were per-
formed serially. Each query updating the member’s last login timestamp needed to
finish before the next began. Ten queries that were considered performant, each fin-
ishing within 50ms, would add up to half a second in database updates alone. All of
these queries would have to finish prior to sending the authorization response and
redirect back to the user. In addition, any operational issues on a database server com-
pounded the problem. If one database server in the chain of servers started respond-
ing slowly or became unresponsive, members could no longer login into the site.
To decouple the user-facing login application from directly writing to the database.
I published messages to Message-oriented-middleware or a centralized message bro-
ker, which would distribute the message to consumer applications that handle the
required database writes. While I first experimented with several different message
brokers, ultimately I landed on RabbitMQ as my broker of choice.

DEFINITION Message-oriented-middleware (MOM) is defined as software or

hardware infrastructure that allows for the sending and receiving of messages
from distributed systems. RabbitMQ fills this role with functionality that pro-
vides advanced routing and message distribution, even with wide-area net-
work (WAN) tolerances to support reliable, distributed systems that
interconnect with other systems easily.

After decoupling the login process from the required database updates, a new level of
freedom was discovered. Members were able to quickly login because we were no lon-
ger updating the database as part of the authentication process. Instead a member
login message was published with all of the information needed to update any data-
base, and consumer applications were written that updated each database table inde-
pendently (figure 1.3). This login message would not contain authentication for the
member, but instead, only the information needed to maintain the member’s last-
login status in our databases and applications. This allowed us to horizontally scale
database writes with more control. By controlling the number of consumer applica-
tions writing for a specific database server, we were able to throttle database writes for
servers that had started to strain under the load created by new site growth, while we
worked through their own unique scaling issues.
As I detail the advantages of a messaging based architecture, it is important to note
that they (they who?) could also impact the performance of systems like the login
architecture described. Any number of problems may impact publisher performance,
from networking issues to RabbitMQ throttling message publishers. When such events
happen, your application will see degraded performance. In addition to the horizon-

Licensed to Marocco Aurélien <[email protected]>

76 CHAPTER 1 Foundational RabbitMQ

Login Application

2. RabbitMQ publishes the

login event message to all of
the consumers who should
receive it.
1. The member authenticates
and a message is published to
RabbitMQ and then the
application redirects the
authenticated member to the
logged-in URL

C C C

DB DB DB

3. Each consumer independently

performs its own database task.

Figure 1.3 After: Using RabbitMQ, loosely coupled data is published to each database asynchronously
and independently, allowing the login application to proceed without waiting on any database writes.

tal scaling of consumers, it is wise to plan for horizontal scaling of message brokers to
allow for better message throughput and publisher performance.

1.3.1 Decoupling your application

The use of messaging-oriented-middleware can provide tremendous advantages for
organizations looking to create flexible application architectures that are data centric.
By moving to a loosely coupled design using RabbitMQ, application architectures are no
longer bound to database write performance and can easily add new applications to act
upon the data without touching any of the core applications. Consider figure 1.4 dem-
onstrating the design of a tightly coupled application communicating with a database:

Figure 1.4 When communicating with a database, a tightly coupled application must wait for the
database server to respond to continue processing.

Licensed to Marocco Aurélien <[email protected]>

 The advantages of loosely coupled architectures 77

1.3.2 Decoupling database writes

In a tightly coupled architecture, the application must wait for the database server to
respond before it can finish a transaction. This design has the potential to create per-
formance bottlenecks in both synchronous and asynchronous applications. If the
database server slows down due to poor tuning or hardware issues, the application will
also become slower. Should the database no longer respond, or it crashes, the applica-
tion will potentially crash as well. By decoupling the database from the application, a
loosely coupled architecture is created. In this architecture, RabbitMQ as messaging-
oriented-middleware acts as an intermediary for the data prior to some action being
taken with it in the database. A consumer application picks up the data from the Rab-
bitMQ server, performing the action with the database (figure 1.5).

Figure 1.5 A loosely coupled application allows the application that would have saved the data directly
in the database to publish the data to RabbitMQ, allowing for the asynchronous processing of data.

In this model, should a database need to be taken offline for maintenance, or should
the write workload become too heavy, you can throttle the consumer application or
stop it. Until the consumer is able to receive the message, the data will persist in the
queue. The ability to pause or throttle consumer application behavior is just one
advantage of using this type of architecture.

1.3.3 Seamlessly adding new functionality

Loosely coupled architectures leveraging RabbitMQ allows data to be repurposed as
well. The data written to a database can also be used for other purposes. RabbitMQ
will handle all of the duplication of the message content and can route it to multiple
consumers for multiple purposes (figure 1.6).

1.3.4 Replication of data and events

Expanding upon this model, RabbitMQ provides built-in tools for cross-datacenter
distribution of data, allowing for federated delivery and synchronization of applica-

Licensed to Marocco Aurélien <[email protected]>

78 CHAPTER 1 Foundational RabbitMQ

Figure 1.6 By using RabbitMQ, the publishing application does not need to be changed in order to
deliver the same data to both a new cloud-based service and the original database.

tions. Federation provides a mechanism that allows RabbitMQ to push messages to

remote RabbitMQ instances, accounting for WAN tolerances and network splits.
Using the RabbitMQ federation plugin, it is easy to add a RabbitMQ server or cluster
to a second data center. This is illustrated in figure 1.7 where the data from the origi-
nal application can now be processed in two different locations over the Internet.

1.3.5 Multi-Master federation of data and events

Expanding upon this concept, by adding the same front-end application to Data Center
#2 and setting the RabbitMQ servers to bi-directionally federate data, you can have
highly available applications in different physical locations. Messages from the applica-
tion in either data center are sent to consumers in both data centers, allowing for
redundancy in data storage and processing (figure 1.8). This approach to application
architecture can provide a scale-out approach to applications, providing geographic
proximity for users and a cost-effective way to distribute your application infrastructure.

NOTE: As with any architecture decision, using messaging-oriented-middle-

ware introduces a degree of operational complexity. Because a message bro-
ker becomes a center point in your application design, a new point of failure
is introduced. There are strategies, which we will cover in this book, to create
highly available solutions to minimize this risk. In addition, adding a message
broker creates a new application to manage. Configuration, server resources,
and monitoring must be taken into account when weighing the tradeoffs of
introducing a message broker to your architecture; I will teach you how to
account for these and other concerns as you proceed through the book.

Licensed to Marocco Aurélien <[email protected]>

 The advantages of loosely coupled architectures 79

Figure 1.7 By leveraging RabbitMQ’s federation plugin, messages can be duplicated to perform
the same work in multiple data centers.

Figure 1.8 Bi-directional federation of data allows for the same data events to be received
processed in both data centers.
Licensed to Marocco Aurélien <[email protected]>
80 CHAPTER 1 Foundational RabbitMQ

1.4 The Advanced Message Queuing Model

Many of RabbitMQ’s strengths and flexibility come from the AMQP specification.
Unlike protocols like HTTP and SMTP, the AMQP specification defines not only a
network protocol, but it also defines server-side services and behaviors. To have a com-
mon way to refer to this information is the Advanced Message Queuing (AMQ)
Model. The AMQ model logically defines three abstract components in broker soft-
ware that determine the routing behavior of messages:
An exchange, the component of the message broker that routes messages to
queues
A queue, a data structure on disk or in memory that stores messages
A binding, which tells the exchange which queue messages should be stored in
The flexibility of RabbitMQ comes from the dynamic nature of how messages can be
routed through exchanges to queues. These bindings between exchanges and queues,
and the message routing dynamics they create, are a foundational component of
implementing messaging based architecture. Creating the right structure by using
these basic tools in RabbitMQ allows your applications to scale and easily change with
the underlying business needs.
EXCHANGES
The first piece of information that RabbitMQ needs in order to route messages to
their proper destination is an exchange to route them through. Exchanges are one of
three components defined by the AMQ model. An exchange receives messages sent
into RabbitMQ and determines where to send them. Exchanges define the routing
behaviors that are applied to messages, usually by examining data attributes passed
along, or that are contained within the message’s properties. RabbitMQ has multiple
exchange types, each with different routing behaviors. In addition, it offers a plug-in
based architecture for custom exchanges. Figure 1.9 shows a logical view of a pub-
lisher sending a message to RabbitMQ, routing a message through an exchange, the
first component of the AMQ model.

Figure 1.9 When a publisher sends a message into RabbitMQ, it first goes to an exchange.

Licensed to Marocco Aurélien <[email protected]>

 The Advanced Message Queuing Model 81

QUEUES
A queue is responsible for storing received messages and may contain configuration
information that defines what it is able to do with a message. Queues may hold mes-
sages in RAM only or it may persist them to disk prior to delivering messages from a
queue in first-in, first-out (FIFO) order.
BINDINGS
To define a relationship between queues and exchanges, the AMQ model defines a
binding. In RabbitMQ, bindings or binding-keys tell an exchange which queues to
deliver messages to. For some exchange types it will also instruct the exchange to filter
which messages it can deliver to a queue. When publishing a message to an exchange,
applications use a routing-key attribute. Sometimes this may be a queue name, at other
times it may be a string that semantically describes the message. When a message is
evaluated by an exchange to be routed to the appropriate queues, the message’s rout-
ing-key is evaluated against the binding-key (figure 1.10). In other words, the binding-
key is the glue that binds a queue to an exchange and the routing-key is the criteria
that is evaluated against it.

Figure 1.10 A queue is bound to an exchange, providing the information the exchange needs to route a
message to it.

In the simplest of scenarios, the routing key may be the queue name, though this
varies with each exchange type. In RabbitMQ, each exchange type is likely to treat
routing-keys in a different way; some exchanges invoke simple equality checks and
others use more complex pattern extractions from the routing-key. There is even an
exchange type that ignores the routing-key outright in favor of other information in
the message properties.
In addition to binding queues to exchanges, as defined in the AMQ model, Rab-
bitMQ extends the AMQP specification to allow exchanges to bind to other

Licensed to Marocco Aurélien <[email protected]>

82 CHAPTER 1 Foundational RabbitMQ

exchanges. This feature creates a great deal of flexibility in creating different routing
patterns for messages. You will learn more about routing patterns available when you
use exchanges, and about exchange-to-exchange bindings in chapter 6, Common
Messaging Patterns.

1.5 Summary
RabbitMQ, as messaging-oriented-middleware, is an exciting technology that enables
operational flexibility that is otherwise difficult to achieve without the loosely coupled
application architecture it enables. By diving deep into RabbitMQ’s AMQP founda-
tion and behaviors, this book is a valuable reference, providing insight into how your
applications can leverage its robust and powerful features. In particular, you will soon
learn how to publish messages, and use the dynamic routing features in RabbitMQ to
selectively sip from the fire hose of data your application can send; data that once may
have been deeply buried in tightly coupled code and processes in your environment.
Whether you are an application developer or a high-level application architect, it is
advantageous to have deep level knowledge about how your applications can benefit
from RabbitMQ’s diverse functionality. Thus far, you have already learned the most
foundational concepts that comprise the Advanced Message Queuing Model. Expand-
ing on these concepts in Part 1 of this book, you will learn about the Advanced Mes-
sage Queuing Protocol and how it defines the core of RabbitMQ’s behavior.
Because this book will be hands-on with the goal of imparting the knowledge
required to use RabbitMQ, the most demanding of environments, you will start work-
ing with code in the next chapter. By learning “how to speak Rabbit” you will leverage
the fundamentals of the Advanced Message Queuing Protocol, writing code to send
and receive messages with RabbitMQ. To speak Rabbit, you will be using a Python
based library called rabbitpy, a library that was written specifically for the code exam-
ples in this book; I’ll introduce it to you in the next chapter. Even if you are an experi-
enced developer who has written applications that communicate with RabbitMQ, you
should at least browse through the next chapter to understand what is happening at
the protocol level when you are using RabbitMQ via the AMQP protocol.

Licensed to Marocco Aurélien <[email protected]>

 Summary 83

Any large application needs an efficient way to handle

the constant messages passing between components in
the system. Billed as "messaging that just works," the
RabbitMQ message broker initially appeals to develop-
ers because it's lightweight, easy to set up, and low main-
tenance. They stick with it, though, because it's
powerful, fast, and up to nearly anything you can throw
at it. This book takes you beyond the basics and explores
the challenges of clustering and distributing messages
across enterprise-level data-centers using RabbitMQ.
RabbitMQ in Depth is a practical guide to building
and maintaining message-based systems. This book
covers detailed architectural and operational use of RabbitMQ with an emphasis on
not just how it works but why it works the way it does. You'll find examples and
detailed explanations of everything from low-level communication to integration with
third-party systems. You'll also find the insights you need to make core architectural
choices and develop procedures for effective operational management.

What's inside
Understanding the AMQP model
Communicating via MQTT, Stomp, and HTTP
Valuable troubleshooting techniques
Integrating with Java technologies like Hadoop and Esper
Database integrations with PostgreSQL and Riak
Written for programmers with a basic understanding of messaging oriented systems
and RabbitMQ.

Licensed to Marocco Aurélien <[email protected]>

 Netty in Action

U sually cloud computing is heavily reliant on networking. Distributed

systems offer scalability and high-availability by distributing workloads on
server fleets. Communication between different applications and other systems
becomes more and more important. But developing applications that are rely-
ing heavily on networking is. The Netty framework makes it easier to create
high-performance networking applications, so we’ve selected this chapter
from Netty in Action, which should give you an idea where Netty could fit in
your own projects.

Licensed to Marocco Aurélien <[email protected]>

 Chapter 14 from Netty in Action by Norman
Maurer and Marvin Allen Wolfthal.

Case studies, part 1

This chapter covers

Droplr
Firebase
Urban Airship

In this chapter we’ll present the first of two sets of case studies contributed by com-
panies that have used Netty extensively in their internal infrastructure. We hope
that these examples of how others have utilized the framework to solve real-world
problems will broaden your understanding of what you can accomplish with Netty.

NOTE The author or authors of each study were directly involved in the
project they discuss.

14.1 Droplr—building mobile services

Bruno de Carvalho, Lead Architect
At Droplr we use Netty at the heart of our infrastructure, in everything from our
API servers to auxiliary services.

85

Licensed to Marocco Aurélien <[email protected]>

86 CHAPTER 14 Case studies, part 1

This is a case study on how we moved from a monolithic and sluggish LAMP1 appli-
cation to a modern, high-performance and horizontally distributed infrastructure,
implemented atop Netty.

14.1.1 How it all started

When I joined the team, we were running a LAMP application that served both as the
front end for users and as an API for the client applications—among which, my
reverse-engineered, third-party Windows client, windroplr.
Windroplr went on to become Droplr for Windows, and I, being mostly an infrastruc-
ture guy, eventually got a new challenge: completely rethink Droplr’s infrastructure.
By then Droplr had established itself as a working concept, so the goals were pretty
standard for a 2.0 version:
Break the monolithic stack into multiple horizontally scalable components
Add redundancy to avoid downtime
Create a clean API for clients
Make it all run on HTTPS
Josh and Levi, the founders, asked me to “make it fast, whatever it takes.”
I knew those words meant more than making it slightly faster or even a lot faster.
“Whatever it takes” meant a full order of magnitude faster. And I knew then that Netty
would eventually play an important role in this endeavor.

14.1.2 How Droplr works

Droplr has an extremely simple workflow: drag a file to the app’s menu bar icon and
Droplr uploads the file. When the upload completes, Droplr copies a short URL to the
file—the drop—to the clipboard.
That’s it. Frictionless, instant sharing.
Behind the scenes, drop metadata is stored in a database—creation date, name,
number of downloads, and so on—and the files are stored on Amazon S3.

14.1.3 Creating a faster upload experience

The upload flow for Droplr’s first version was woefully naïve:
1 Receive upload
2 Upload to S3
3 Create thumbnails if it’s an image
4 Reply to client applications
A closer look at this flow quickly reveals two choke points on steps 2 and 3. No matter
how fast the upload from the client to our servers, the creation of a drop would always
go through an annoying hiatus after the actual upload completed, until the successful

1
An acronym for a typical application technology stack; originally Linux, Apache Web Server, MySQL, and
PHP.

Licensed to Marocco Aurélien <[email protected]>

 Droplr—building mobile services 87

response was received—because the file would still need to be uploaded to S3 and
have its thumbnails generated.
The larger the file, the longer the hiatus. For very large files the connection would
eventually time out waiting for the okay from the server. Back then Droplr could offer
uploads of only up to 32 MB per file because of this very problem.
There were two distinct approaches to cut down upload times:
Approach A, optimistic and apparently simpler (figure 14.1):
– Fully receive the file
– Save to the local filesystem and immediately return success to client
– Schedule an upload to S3 some time in the future

Post

Local

Put

Figure 14.1 Approach A,

S3 optimistic and apparently simpler

Approach B, safe but complex (figure 14.2):

– Pipe the upload from the client directly to S3, in real time (streaming)

Post

Put

Figure 14.2 Approach B, safe

S3 but complex

THE OPTIMISTIC AND APPARENTLY SIMPLER APPROACH

Returning a short URL after receiving the file creates an expectation—one could even
go as far as calling it an implicit contract—that the file is immediately available at that
URL. But there is no guarantee that the second stage of the upload (actually pushing

Licensed to Marocco Aurélien <[email protected]>

88 CHAPTER 14 Case studies, part 1

the file to S3) will ultimately succeed, and the user could end up with a broken link
that might get posted on Twitter or sent to an important client. This is unacceptable,
even if it happens on one in every hundred thousand uploads.
Our current numbers show that we have an upload failure rate slightly below
0.01% (1 in every 10,000), the vast majority being connection timeouts between client
and server before the upload actually completes.
We could try to work around it by serving the file from the machine that received it
until it is finally pushed to S3, but this approach is in itself a can of worms:
If the machine fails before a batch of files is completely uploaded to S3, the files
would be forever lost.
There would be synchronization issues across the cluster (“Where is the file for
this drop?”).
Extra, complex logic would be required to deal with edge cases, and this keeps
creating more edge cases.
Thinking through all the pitfalls with every workaround, I quickly realized that it’s a
classic hydra problem—for each head you chop off, two more appear in its place.
THE SAFE BUT COMPLEX APPROACH
The other option required low-level control over the whole process. In essence, we
had to be able to
Open a connection to S3 while receiving the upload from the client.
Pipe data from the client connection to the S3 connection.
Buffer and throttle both connections:
– Buffering is required to keep a steady flow between both client-to-server and
server-to-S3 legs of the upload.
– Throttling is required to prevent explosive memory consumption in case the
server-to-S3 leg of the upload becomes slower than the client-to-server leg.
Cleanly roll everything back on both ends if things went wrong.
It seems conceptually simple, but it’s hardly something your average webserver can
offer. Especially when you consider that in order to throttle a TCP connection, you
need low-level access to its socket.
It also introduced a new challenge that would ultimately end up shaping our final
architecture: deferred thumbnail creation.
This meant that whichever technology stack the platform ended up being built
upon, it had to offer not only a few basic things like incredible performance and sta-
bility but also the flexibility to go bare metal (read: down to the bytes) if required.

14.1.4 The technology stack

When kick-starting a new project for a webserver, you’ll end up asking yourself, “Okay,
so what frameworks are the cool kids using these days?” I did too.

Licensed to Marocco Aurélien <[email protected]>

 Droplr—building mobile services 89

Going with Netty wasn’t a no-brainer; I explored plenty of frameworks, having in

mind three factors that I considered to be paramount:
It had to be fast. I wasn’t about to replace a low-performance stack with another
low-performance stack.
It had to scale. Whether it had 1 or 10,000 connections, each server instance
would have to be able to sustain throughput without crashing or leaking mem-
ory over time.
It had to offer low-level data control. Byte-level reads, TCP congestion control,
the works.
Factors 1 and 2 pretty much excluded any noncompiled language. I’m a sucker for
Ruby and love lightweight frameworks like Sinatra and Padrino, but I knew the kind
of performance I was looking for couldn’t be achieved by building on these blocks.
Factor 2, on its own, meant that whatever the solution, it couldn’t rely on blocking
I/O. By this point in the book, you certainly understand why non-blocking I/O was
the only option.
Factor 3 was trickier. It meant finding the perfect balance between a framework
that would offer low-level control of the data it received, but at the same time would
be fast to develop with and build upon. This is where language, documentation, com-
munity, and other success stories come into play.
At this point I had a strong feeling Netty was my weapon of choice.
THE BASICS: A SERVER AND A PIPELINE
The server is merely a ServerBootstrap built with an NioServerSocketChannel-
Factory, configured with a few common handlers and an HTTP RequestController
at the end, as shown here.

Listing 14.1 Setting up the ChannelPipeline

pipelineFactory = new ChannelPipelineFactory() { IdleStateHandler

@Override shuts down inactive
public ChannelPipeline getPipeline() throws Exception { connections
ChannelPipeline pipeline = Channels.pipeline();
pipeline.addLast("idleStateHandler", new IdleStateHandler(...));
pipeline.addLast("httpServerCodec", new HttpServerCodec());
pipeline.addLast("requestController",
new RequestController(...)); HttpServerCodec converts
return pipeline; incoming bytes to
} Adds a HttpRequests and outgoing
}; RequestController HttpResponses to bytes
to the pipeline
The RequestController is the only custom Droplr code in the pipeline and is proba-
bly the most complex part of the whole webserver. Its role is to handle initial request
validations and, if all is well, route the request to the appropriate request handler. A
new instance is created for every established client connection and lives for as long as
that connection remains active.

Licensed to Marocco Aurélien <[email protected]>

90 CHAPTER 14 Case studies, part 1

The request controller is responsible for

Handling load peaks

HTTP pipeline management

Setting up a context for request handling
Spawning new request handlers

Feeding request handlers

Handling internal and external errors

Here is a quick rundown of the relevant parts of the RequestController.

Listing 14.2 The RequestController

public class RequestController

extends IdleStateAwareChannelUpstreamHandler {

@Override
public void channelIdle(ChannelHandlerContext ctx,
IdleStateEvent e) throws Exception {
// Shut down connection to client and roll everything back.
}

@Override public void channelConnected(ChannelHandlerContext ctx,

ChannelStateEvent e) throws Exception {
if (!acquireConnectionSlot()) {
// Maximum number of allowed server connections reached,
// respond with 503 service unavailable
// and shutdown connection.
} else {
// Set up the connection's request pipeline.
}
}

@Override public void messageReceived(ChannelHandlerContext ctx,

MessageEvent e) throws Exception { The gist of
if (isDone()) return; Droplr’s
if (e.getMessage() instanceof HttpRequest) { server request
handleHttpRequest((HttpRequest) e.getMessage()); validation
} else if (e.getMessage() instanceof HttpChunk) {
handleHttpChunk((HttpChunk)e.getMessage());
}
} If there’s an active handler for the
} current request and it accepts
chunks, it then passes on the chunk.

As explained previously in this book, you should never execute non-CPU-bound code
on Netty’s I/O threads—you’ll be stealing away precious resources from Netty and
thus affecting the server’s throughput.
For this reason, both the HttpRequest and HttpChunk may hand off the execution
to the request handler by switching over to a different thread. This happens when the

Licensed to Marocco Aurélien <[email protected]>

 Droplr—building mobile services 91

request handlers aren’t CPU-bound, whether because they access the database or per-
form logic that’s not confined to local memory or CPU.
When thread-switching occurs, it’s imperative that all the blocks of code execute in
serial fashion; otherwise we’d risk, for an upload, having HttpChunk n-1 being pro-
cessed after HttpChunk n and thus corrupting the body of the file. (We’d be swapping
how bytes were laid out in the uploaded file.) To cope with this, I created a custom
thread-pool executor that ensures all tasks sharing a common identifier will be exe-
cuted serially.
From here on, the data (requests and chunks) ventures out of the realms of Netty
and Droplr.
I’ll explain briefly how the request handlers are built for the sake of shedding
some light on the bridge between the RequestController—which lives in Netty-
land—and the handlers—Droplr-land. Who knows, maybe this will help you architect
your own server!
THE REQUEST HANDLERS
Request handlers provide Droplr’s functionality. They’re the endpoints behind URIs
such as /account or /drops. They’re the logic cores—the server’s interpreters of cli-
ents’ requests.
Request handler implementations are where the framework actually becomes
Droplr’s API server.
THE PARENT INTERFACE
Each request handler, whether directly or through a subclass hierarchy, is a realization
of the interface RequestHandler.
In its essence, the RequestHandler interface represents a stateless handler for
requests (instances of HttpRequest) and chunks (instances of HttpChunk). It’s an
extremely simple interface with a couple of methods to help the request controller
perform and/or decide how to perform its duties, such as:
Is the request handler stateful or stateless? Does it need to be cloned from a
prototype or can the prototype be used to handle the request?
Is the request handler CPU or non-CPU bound? Can it execute on Netty’s
worker threads or should it be executed in a separate thread pool?
Roll back current changes.
Clean up any used resources.
This interface is all the RequestController knows about actions. Through its very
clear and concise interface, the controller can interact with stateful and stateless, CPU-
bound and non-CPU-bound handlers (or combinations of these) in an isolated and
implementation-agnostic fashion.
HANDLER IMPLEMENTATIONS
The simplest realization of RequestHandler is AbstractRequestHandler, which repre-
sents the root of a subclass hierarchy that becomes ever more specific until it reaches

Licensed to Marocco Aurélien <[email protected]>

92 CHAPTER 14 Case studies, part 1

the actual handlers that provide all of Droplr’s functionality. Eventually it leads to the
stateful implementation SimpleHandler, which executes in a non-IO-worker thread
and is therefore not CPU-bound. SimpleHandler is ideal for quickly implementing
endpoints that do the typical tasks of reading in JSON, hitting the database, and then
writing out some JSON.
THE UPLOAD REQUEST HANDLER
The upload request handler is the crux of the whole Droplr API server. It was the action
that shaped the design of the webserver module—the frameworky part of the server—
and it’s by far the most complex and tuned piece of code in the whole stack.
During uploads, the server has dual behaviors:
On one side, it acts as a server for the API clients that are uploading the files.
On the other side, it acts as client to S3 to push the data it receives from the API
clients.
To act as a client, the server uses an HTTP client library that is also built with Netty.1
This asynchronous library exposes an interface that perfectly matches the needs of
the server. It begins executing an HTTP request and allows data to be fed to it as it
becomes available, and this greatly reduces the complexity of the client facade of the
upload request handler.

14.1.5 Performance
After the initial version of the server was complete, I ran a batch of performance tests.
The results were nothing short of mind blowing. After continuously increasing the
load in disbelief, I saw the new server peak at 10~12x faster uploads over the old LAMP
stack—a full order of magnitude faster—and it could handle over 1000x more concur-
rent uploads, for a total of nearly 10 k concurrent uploads (running on a single EC2
large instance).
The following factors contributed to this:
It was running in a tuned JVM.
It was running in a highly tuned custom stack, created specifically to address
this problem, instead of an all-purpose web framework.
The custom stack was built with Netty using NIO (selector-based model), which
meant it could scale to tens or even hundreds of thousands of concurrent con-
nections, unlike the one-process-per-client LAMP stack.
There was no longer the overhead of receiving a full file and then uploading it
to S3 in two separate phases. The file was now streamed directly to S3.

1
You can find the HTTP client library at https://github.com/brunodecarvalho/http-client.

Licensed to Marocco Aurélien <[email protected]>

 Firebase—a real-time data synchronization service 93

Because the server was now streaming files,

– It was not spending time on I/O operations, writing to temporary files and
later reading them in the second stage of the upload.
– It was using less memory for each upload, which meant more parallel uploads
could take place.
Thumbnail generation became an asynchronous post-process.

14.1.6 Summary—standing on the shoulders of giants

All of this was possible thanks to Netty’s incredibly well-designed API and performant
nonblocking I/O architecture.
Since the launch of Droplr 2.0 in December 2011, we’ve had virtually zero down-
time at the API level. A couple of months ago we interrupted a year-and-a-half clean
run of 100% infrastructure uptime due to a scheduled full-stack upgrade (databases,
OS, major server and daemons codebase upgrade) that took just under an hour.
The servers soldier on, day after day, taking hundreds—sometimes thousands—of
concurrent requests per second, all the while keeping both memory and CPU use to
levels so low it’s hard to believe they’re actually doing such an incredible amount of
work:
CPU use rarely ever goes above 5%.
Memory footprint can’t be accurately described as the process starts with 1 GB
of preallocated memory, with the JVM configured to grow up to 2 GB if neces-
sary, and not a single time in the past two years has this happened.
Anyone can throw more machines at any given problem, but Netty helped Droplr
scale intelligently, and keep the server bills pretty low.

14.2 Firebase—a real-time data synchronization service

Sara Robinson, VP of Developer Happiness
Greg Soltis, VP of Cloud Architecture

Real-time updates are an integral part of the user experience in modern applications.
As users come to expect this behavior, more and more applications are pushing data
changes to users in real time. Real-time data synchronization is difficult to achieve
with the traditional three-tiered architecture, which requires developers to manage
their own ops, servers, and scaling. By maintaining real-time, bidirectional communi-
cation with the client, Firebase provides an immediately intuitive experience allowing
developers to synchronize application data across diverse clients in a few minutes—all
without any backend work, servers, ops, or scaling required.
Implementing this presented a difficult technical challenge, and Netty was the
optimal solution in building the underlying framework for all network communica-
tions in Firebase. This study will provide an overview of Firebase’s architecture, and

Licensed to Marocco Aurélien <[email protected]>

94 CHAPTER 14 Case studies, part 1

then examine three ways Firebase uses Netty to power its real-time synchronization
service:
Long polling
HTTP 1.1 keep-alive and pipelining
Control of SSL handler

14.2.1 The Firebase architecture

Firebase allows developers to get an application up and running using a two-tiered
architecture. Developers simply include the Firebase library and write client-side
code. The data is exposed to the developer’s code as JSON and is cached locally. The
library handles synchronizing this local cache with the master copy, which is stored
on Firebase’s servers. Changes made to any data are synchronized in real time to
potentially hundreds of thousands of clients connected to Firebase. The interaction
between multiple clients across platforms and devices and Firebase is depicted in fig-
ure 14.3.

Figure 14.3 Firebase architecture

Firebase servers take incoming data updates and immediately synchronize them to all
of the connected clients that have registered interest in the changed data. To enable
real-time notification of state changes, clients maintain an active connection to Fire-
base at all times. This connection may range from an abstraction over a single Netty
channel to an abstraction over multiple channels or even multiple, concurrent
abstractions if the client is in the middle of switching transport types.
Because clients can connect to Firebase in a variety of ways, it’s important to keep
the connection code modular. Netty’s Channel abstraction is a fantastic building block
for integrating new transports into Firebase. In addition, the pipeline-and-handler
pattern makes it simple to keep transport-specific details isolated and provide a com-
mon message stream abstraction to the application code. Similarly, this greatly simpli-
fies adding support for new protocols. Firebase added support for a binary transport
simply by adding a few new handlers to the pipeline. Netty’s speed, level of abstrac-
tion, and fine-grained control made it an excellent framework for implementing real-
time connections between the client and server.

Licensed to Marocco Aurélien <[email protected]>

 Firebase—a real-time data synchronization service 95

14.2.2 Long polling

Firebase uses both long polling and WebSocket transports. The long-polling transport
is highly reliable across all browsers, networks, and carriers; the WebSocket-based
transport is faster but not always available due to limitations of browsers/clients. Ini-
tially, Firebase connects using long polling and then upgrades to WebSockets if possi-
ble. For the minority of Firebase traffic that doesn’t support WebSockets, Firebase
uses Netty to implement a custom library for long polling tuned to be highly perfor-
mant and responsive.
The Firebase library logic deals with bidirectional streams of messages with notifica-
tions when either side closes the stream. Although this is relatively simple to implement
on top of TCP or WebSockets, it presents a challenge when dealing with a long-polling
transport. The two properties that must be enforced for the long-polling case are
Guaranteed in-order delivery of messages
Close notifications
GUARANTEED IN-ORDER DELIVERY OF MESSAGES
In-order delivery for long polling can be achieved by having only a single request out-
standing at a given time. Because the client won’t send another request until it
receives a response from its last request, it can guarantee that its previous messages
were received and that it’s safe to send more. Similarly, on the server side, there won’t
be a new request outstanding until the client has received the previous response.
Therefore, it’s always safe to send everything that’s buffered up in between requests.
However, this leads to a major drawback. Using the single-request technique, both the
client and server spend a significant amount of time buffering up messages. If the cli-
ent has new data to send but already has an outstanding request, for example, it must
wait for the server to respond before sending the new request. This could take a long
time if there’s no data available on the server.
A more performant solution is to tolerate more requests being in flight concur-
rently. In practice, this can be achieved by swapping the single-request pattern for the
at-most-two-requests pattern. This algorithm has two parts:
Whenever a client has new data to send, it sends a new request unless two are
already in flight.
Whenever the server receives a request from a client, if it already has an open
request from the client, it immediately responds to the first even if there is
no data.
This provides an important improvement over the single-request pattern: both the cli-
ent’s and server’s buffer time are bound to at most a single network round-trip.
Of course, this increase in performance doesn’t come without a price; it results in
a commensurate increase in code complexity. The long-polling algorithm no longer
guarantees in-order delivery, but a few ideas from TCP can ensure that messages are
delivered in order. Each request sent by the client includes a serial number, incre-

Licensed to Marocco Aurélien <[email protected]>

96 CHAPTER 14 Case studies, part 1

mented for each request. In addition, each request includes metadata about the
number of messages in the payload. If a message spans multiple requests, the portion
of the message contained in this payload is included in the metadata.
The server maintains a ring buffer of incoming message segments and processes
them as soon as they’re complete and no incomplete messages are ahead of them.
Downstream is easier because the long-polling transport responds to an HTTP GET
request and doesn’t have the same restrictions on payload size. In this case, a serial
number is included and is incremented once for each response. The client can pro-
cess all messages in the list as long as it has received all responses up to the given serial
number. If it hasn’t, it buffers the list until it receives the outstanding responses.
CLOSE NOTIFICATIONS
The second property enforced in the long-polling transport is close notification. In
this case, having the server be aware that the transport has closed is significantly more
important than having the client recognize the close. The Firebase library used by cli-
ents queues up operations to be run when a disconnect occurs, and those operations
can have an impact on other still-connected clients. So it’s important to know when a
client has actually gone away. Implementing a server-initiated close is relatively simple
and can be achieved by responding to the next request with a special protocol-level
close message.
Implementing client-side close notifications is trickier. The same close notification
can be used, but there are two things that can cause this to fail: the user can close the
browser tab, or the network connection could disappear. The tab-closure case is han-
dled with an iframe that fires a request containing the close message on page unload.
The second case is dealt with via a server-side timeout. It’s important to pick your
timeout values carefully, because the server is unable to distinguish a slow network
from a disconnected client. That is to say, there’s no way for the server to know that a
request was actually delayed for a minute, rather than the client losing its network
connection. It’s important to choose an appropriate timeout that balances the cost of
false positives (closing transports for clients on slow networks) against how quickly the
application needs to be aware of disconnected clients.
Figure 14.4 demonstrates how the Firebase long-polling transport handles differ-
ent types of requests.
In this diagram, each long-poll request indicates different types of scenarios. Ini-
tially, the client sends a poll (poll 0) to the server. Some time later, the server receives
data from elsewhere in the system that is destined for this client, so it responds to poll
0 with the data. As soon as the poll returns, the client sends a new poll (poll 1),
because it currently has none outstanding. A short time later, the client needs to send
data to the server. Since it only has a single poll outstanding, it sends a new one (poll 2)
that includes the data to be delivered. Per the protocol, as soon as the server has two
simultaneous polls from the same client, it responds to the first one. In this case, the
server has no data available for the client, so it sends back an empty response. The cli-
ent also maintains a timeout and will send a second poll when it fires, even if it has no

Licensed to Marocco Aurélien <[email protected]>

 Firebase—a real-time data synchronization service 97

Client Server

poll 0

Data for client is available

nse 0
Respo e data
y in g th
(Ca rr
th e server)
fro m

poll 1
poll 2
(Co
data s ntaining
ent fr
the cli om
ent Data from client is sent
for processing

)
o data
n se 1 (N
Respo Figure 14.4 Long polling

additional data to send. This insulates the system from failures due to browsers timing
out slow requests.

14.2.3 HTTP 1.1 keep-alive and pipelining

With HTTP 1.1 keep-alive, multiple requests can be sent on one connection to a server.
This allows for pipelining—new requests can be sent without waiting for a response
from the server. Implementing support for pipelining and keep-alive is typically
straightforward, but it gets significantly more complex when mixed with long polling.
If a long-polling request is immediately fol-
lowed by a REST (Representational State Trans-
Client Server
fer) request, there are some considerations that long p
oll 0
need to be taken into account to ensure the
browser performs properly. A channel may mix
asynchronous messages (long-poll requests) with REST

synchronous messages (REST requests). When a

synchronous request comes in on one channel,
long p
Firebase must synchronously respond to all pre- oll 1

ceding requests in that channel in order. For exam-

ple, if there’s an outstanding long-poll request, the
long-polling transport needs to respond with a no- oll 0
long p
op before handling the REST request.
Figure 14.5 illustrates how Netty lets Firebase
respond to multiple request types in one socket. REST
If the browser has more than one connection
open and is using long polling, it will reuse the
connection for messages from both of those open Figure 14.5 Network diagram

Licensed to Marocco Aurélien <[email protected]>

98 CHAPTER 14 Case studies, part 1

tabs. Given long-polling requests, this is difficult and requires proper management of
a queue of HTTP requests. Long-polling requests can be interrupted, but proxied
requests can’t. Netty made serving multiple request types easy:
Static HTML pages —Cached content that can be returned with no processing;
examples include a single-page HTML app, robots.txt, and crossdomain.xml.
REST requests —Firebase supports traditional GET, POST, PUT, DELETE, PATCH, and
OPTIONS requests.
WebSocket —A bidirectional connection between a browser and a Firebase server
with its own framing protocol.
Long polling —These are similar to HTTP GET requests but are treated differently
by the application.
Proxied requests —Some requests can’t be handled by the server that receives
them. In that case, Firebase proxies the request to the correct server in its clus-
ter, so that end users don’t have to worry about where data is located. These are
like the REST requests, but the proxying server treats them differently.
Raw bytes over SSL—A simple TCP socket running Firebase’s own framing proto-
col and optimized handshaking.
Firebase uses Netty to set up its pipeline to decode an incoming request and then
reconfigure the remainder of the pipeline appropriately. In some cases, like WebSockets
and raw bytes, once a particular type of request has been assigned a channel, it will
stay that way for its entire duration. In other cases, like the various HTTP requests, the
assignment must be made on a per-message basis. The same channel could handle
REST requests, long-polling requests, and proxied requests.

14.2.4 Control of SslHandler

Netty’s SslHandler class is an example of how Firebase uses Netty for fine-grained
control of its network communications. When a traditional web stack uses an HTTP
server like Apache or Nginx to pass requests to the app, incoming SSL requests have
already been decoded when they’re received by the application code. With a multiten-
ant architecture, it’s difficult to assign portions of the encrypted traffic to the tenant
of the application using a specific service. This is complicated by the fact that multiple
applications could use the same encrypted channel to talk to Firebase (for instance,
the user might have two Firebase applications open in different tabs). To solve this,
Firebase needs enough control in handling SSL requests before they are decoded.
Firebase charges customers based on bandwidth. However, the account to be
charged for a message is typically not available before the SSL decryption has been
performed, because it’s contained in the encrypted payload. Netty allows Firebase to
intercept traffic at multiple points in the pipeline, so the counting of bytes can start as
soon as byes come in off the wire. After the message has been decrypted and processed
by Firebase’s server-side logic, the byte count can be assigned to the appropriate account.
In building this feature, Netty provided control for handling network communica-

Licensed to Marocco Aurélien <[email protected]>

 Firebase—a real-time data synchronization service 99

tions at every layer of the protocol stack, and also allowed for very accurate billing,
throttling, and rate limiting, all of which had significant business implications.
Netty made it possible to intercept all inbound and outbound messages and to
count bytes with a small amount of Scala code.

Listing 14.3 Setting up the ChannelPipeline

case class NamespaceTag(namespace: String)


class NamespaceBandwidthHandler extends ChannelDuplexHandler {

private var rxBytes: Long = 0
private var txBytes: Long = 0
private var nsStats: Option[NamespaceStats] = None

override def channelRead(ctx: ChannelHandlerContext, msg: Object) {

msg match {
case buf: ByteBuf => {

rxBytes += buf.readableBytes( When a message
tryFlush(ctx) comes in, counts the
} number of bytes
case _ => { }
}
super.channelRead(ctx, msg)
}

override def write(ctx: ChannelHandlerContext, msg: Object,

promise: ChannelPromise) {
msg match {

case buf: ByteBuf => { When there is an
txBytes += buf.readableBytes() outbound message,
tryFlush(ctx) counts those bytes as well
super.write(ctx, msg, promise)
}

case tag: NamespaceTag => { If a tag is received, ties
updateTag(tag.namespace, ctx) this channel to an
} account, remembers the
case _ => {
 account, and assigns the
super.write(ctx, msg, promise) current byte counts to it.
}
}
}

private def tryFlush(ctx: ChannelHandlerContext) {

nsStats match {
case Some(stats: NamespaceStats) => { If there’s already a tag
stats.logOutgoingBytes(txBytes.toInt) for the namespace the
txBytes = 0
 channel belongs to,
stats.logIncomingBytes(rxBytes.toInt) assigns the bytes to
rxBytes = 0 that account and
} resets the counters
case None => {

// no-op, we don't have a namespace
}
}
}

Licensed to Marocco Aurélien <[email protected]>

100 CHAPTER 14 Case studies, part 1

private def updateTag(ns: String, ctx: ChannelHandlerContext) {


val (_, isLocalNamespace) = NamespaceOwnershipManager.getOwner(ns)
if (isLocalNamespace) {
nsStats = NamespaceStatsListManager.get(ns)
tryFlush(ctx)
} else {
// Non-local namespace, just flush the bytes
txBytes = 0 If the count isn’t
rxBytes = 0 applicable to this
} machine, ignores it and
} resets the counters
}

14.2.5 Firebase summary

Netty plays an indispensable role in the server architecture of Firebase’s real-time data
synchronization service. It allows support for a heterogeneous client ecosystem, which
includes a variety of browsers, along with clients that are completely controlled by
Firebase. With Netty, Firebase can handle tens of thousands of messages per second
on each server. Netty is especially awesome for several reasons:
It’s fast. It took only a few days to develop a prototype, and was never a produc-
tion bottleneck.
It’s positioned well in the abstraction layer. Netty provides fine-grained control
where necessary and allows for customization at every step of the control flow.
It supports multiple protocols over the same port. HTTP, WebSockets, long polling,
and standalone TCP.
Its GitHub repo is top-notch. Well-written javadocs make it frictionless to develop
against.
It has a highly active community. The community is very responsive on issue main-
tenance and seriously considers all feedback and pull requests. In addition, the
team provides great and up-to-date example code. Netty is an excellent, well-
maintained framework and it has been essential in building and scaling Fire-
base’s infrastructure. Real-time data synchronization in Firebase wouldn’t be
possible without Netty’s speed, control, abstraction, and extraordinary team.

14.3 Urban Airship—building mobile services

Erik Onnen, Vice President of Architecture

As smartphone use grows across the globe at unprecedented rates, a number of ser-
vice providers have emerged to assist developers and marketers toward the end of
providing amazing end-user experiences. Unlike their feature phone predecessors,
smartphones crave IP connectivity and seek it across a number of channels (3G, 4G,
WiFi, WiMAX, and Bluetooth). As more and more of these devices access public net-
works via IP-based protocols, the challenges of scale, latency, and throughput become
more and more daunting for back-end service providers.

Licensed to Marocco Aurélien <[email protected]>

 Urban Airship—building mobile services 101

Thankfully, Netty is well suited to many of the concerns faced by this thundering
herd of always-connected mobile devices. This chapter will detail several practical appli-
cations of Netty in scaling a mobile developer and marketer platform, Urban Airship.

14.3.1 Basics of mobile messaging

Although marketers have long used SMS as a channel to reach mobile devices, a more
recent functionality called push notifications is rapidly becoming the preferred mecha-
nism for messaging smartphones. Push notifications commonly use the less expensive
data channel and the price per message is a fraction of the cost of SMS. The through-
put of push notifications is commonly two to three orders of magnitude higher than
SMS, making it an ideal channel for breaking news. Most importantly, push notifica-
tions give users device-driven control of the channel. If a user dislikes the messaging
from an application, the user can disable notifications for an application or outright
delete the application.
At a very high level, the interaction between a device and push notification behav-
ior is similar to the depiction in figure 14.6.

Public
network API

Core
services

Device
messaging 3rd party Figure 14.6 High-level
channel adapter mobile messaging platform
integration

At a high level, when an application developer wants to send push notifications to a

device, the developer must plan to store information about the device and its applica-
tion installation.1 Commonly, an application installation will execute code to retrieve
a platform-specific identifier and report that identifier back to a centralized service
where the identifier is persisted. Later, logic external to the application installation
will initiate a request to deliver a message to the device.
Once an application installation has registered its identifier with a back-end ser-
vice, the delivery of a push message can in turn take two paths. In the first path, a
message can be delivered directly to the application itself, with the application main-
taining a direct connection to a back-end service. In the second and more common

1
Some mobile OSes allow a form of push notifications called local notifications that would not follow this
approach.

Licensed to Marocco Aurélien <[email protected]>

102 CHAPTER 14 Case studies, part 1

approach, an application will rely on a third party to deliver the message to the appli-
cation on behalf of a back-end service. At Urban Airship, both approaches to deliver-
ing push notifications are used, and both leverage Netty extensively.

14.3.2 Third-party delivery

In the case of third-party push delivery, every push notification platform provides a
different API for developers to deliver messages to application installations. These
APIs differ in terms of their protocol (binary vs. text), authentication (OAuth, X.509,
and so on), and capabilities. Each approach has its own unique challenges for integra-
tion as well as for achieving optimal throughput.
Despite the fact that the fundamental purpose of each of these providers is to
deliver a notification to an application, each takes a different approach with signifi-
cant implications for system integrators. For example, Apple Push Notification Service
(APNS) defines a strictly binary protocol; other providers base their service on some
form of HTTP, all with subtle variations that affect how to best achieve maximum
throughput. Thankfully, Netty is an amazingly flexible tool and it significantly helps
smoothing over the differences between the various protocols.
The following sections will provide examples of how Urban Airship uses Netty to
integrate with two of the listed providers.

14.3.3 Binary protocol example

Apple’s APNS is a binary protocol with a specific, network byte-ordered payload. Send-
ing an APNS notification involves the following sequence of events:
1 Connect a TCP socket to APNS servers over an SSLv3 connection, authenticated
with an X.509 certificate.
2 Format a binary representation of a push message structured according to the
format defined by Apple.1
3 Write the message to the socket.
4 Read from the socket if you’re ready to determine any error codes associated
with a sent message.
5 In the case of an error, reconnect the socket and continue from step 2.
As part of formatting the binary message, the producer of the message is required to
generate an identifier that’s opaque to the APNS system. In the event of an invalid
message (incorrect formatting, size, or device information, for example), the identi-
fier will be returned to the client in the error response message of step 4.

1
For information on APNS: http://docs.aws.amazon.com/sns/latest/dg/mobile-push-apns.html,
http://bit.ly/189mmpG.

Licensed to Marocco Aurélien <[email protected]>

 Urban Airship—building mobile services 103

At face value the protocol seems straightforward, but there are nuances to success-
fully addressing all of the preceding concerns, in particular on the JVM:
The APNS specification dictates that certain payload values should be sent in
big-endian ordering (for example, token length).
Step 3 in the previous sequence requires one of two solutions. Because the JVM
will not allow reading from a closed socket even if data exists in the output buf-
fer, you have two options:
– After a write, perform a blocking read with a timeout on the socket. This has
multiple disadvantages:
– The amount of time to block waiting for an error is non-deterministic. An
error may occur in milliseconds or seconds.
– As socket objects can’t be shared across multiple threads, writes to the
socket must immediately block while waiting for errors. This has dramatic
implications for throughput. If a single message is delivered in a socket
write, no additional messages can go out on that socket until the read
timeout has occurred. When you’re delivering tens of millions of mes-
sages, a three-second delay between messages isn’t acceptable.
– Relying on a socket timeout is an expensive operation. It results in an
exception being thrown and several unnecessary system calls.
– Use asynchronous I/O. In this model, neither reads nor writes block. This
allows writers to continue sending messages to APNS while at the same time
allowing the OS to inform user code when data is ready to be read.
Netty makes addressing all of these concerns trivial while at the same time delivering
amazing throughput.
First, let’s see how Netty simplifies packing a binary APNS message with correct
endian ordering.

Listing 14.4 ApnsMessage implementation

public final class ApnsMessage {

private static final byte COMMAND = (byte) 1; An APNS message always
public ByteBuf toBuffer() { starts with a command 1
short size = (short) (1 + // Command byte in size, so that value
4 + // Identifier is coded as a constant.
4 + // Expiry
2 + // DT length header
Messages size varies,
32 + //DS length
so for efficiency it is
2 + // body length header
calculated before the
body.length); ByteBuf is created.
ByteBuf buf = Unpooled.buffer(size).order(ByteOrder.BIG_ENDIAN);
buf.writeByte(COMMAND);
buf.writeInt(identifier); Various values are At creation the ByteBuf
buf.writeInt(expiryTime); inserted into the buffer is sized exactly and the
from state maintained endianness for APNS is
elsewhere in the class. specified.

Licensed to Marocco Aurélien <[email protected]>

104 CHAPTER 14 Case studies, part 1

buf.writeShort((short) deviceToken.length); The deviceToken

buf.writeBytes(deviceToken); field in this class
buf.writeShort((short) body.length); (not shown) is a
buf.writeBytes(body);
return buf;
b Java byte[].
When the buffer is
} ready, it is simply
}
c returned.

Some important notes on the implementation:

B The length property of a Java array is always an integer. However, the APNS proto-
col requires a 2-byte value. In this case, the length of the payload has been vali-
dated elsewhere, so casting to a short is safe at this location. Note that without
explicitly constructing the ByteBuf to be big endian, subtle bugs could occur
with values of types short and int.
c Unlike the standard java.nio.ByteBuffer, it’s not necessary to flip the buffer
and worry about its position—Netty’s ByteBuf handles read and write position
management automatically.
In a small amount of code, Netty has made trivial the act of creating a properly for-
matted APNS message. Because this message is now packed into a ByteBuf, it can easily
be written directly to a Channel connected to APNS when the message is ready for
sending.
Connecting to APNS can be accomplished via multiple mechanisms, but at its most
basic, a ChannelInitializer that populates the ChannelPipeline with an SslHandler
and a decoder is required.

Listing 14.5 Setting up the ChannelPipeline

public final class ApnsClientPipelineInitializer

An X.509
extends ChannelInitializer<Channel> {
authenticated
private final SSLEngine clientEngine;
request requires a
public ApnsClientPipelineFactory(SSLEngine engine) { javax.net.ssl.SSL-
this.clientEngine = engine; Engine instance.
}
APNS will attempt
@Override to renegotiate
Constructs public void initChannel(Channel channel) throws Exception { SSL shortly after
a Netty final ChannelPipeline pipeline = channel.pipeline(); connection,
SslHandler final SslHandler handler = new SslHandler(clientEngine); need to allow
handler.setEnableRenegotiation(true); renegotiation.
pipeline.addLast("ssl", handler);
pipeline.addLast("decoder", new ApnsResponseDecoder());
}
} This class extends Netty’s
ByteToMessageDecoder and handles cases where
APNS returns an error code and disconnects.

Licensed to Marocco Aurélien <[email protected]>

 Urban Airship—building mobile services 105

It’s worth noting how easy Netty makes negotiating an X.509 authenticated connec-
tion in conjunction with asynchronous I/O. In early prototypes of APNS code at Urban
Airship without Netty, negotiating an asynchronous X.509 authenticated connection
required over 80 lines of code and a thread pool simply to connect. Netty hides all the
complexity of the SSL handshake, the authentication, and most importantly the encryp-
tion of cleartext bytes to cipher text and the key renegotiation that comes along with
using SSL. These incredibly tedious, error prone, and poorly documented APIs in the
JDK are hidden behind three lines of Netty code.
At Urban Airship, Netty plays a role in all connectivity to numerous third-party
push notification services including APNS and Google’s GCM. In every case, Netty is
flexible enough to allow explicit control over exactly how integration takes place from
higher-level HTTP connectivity behavior down to basic socket-level settings such as
TCP keep-alive and socket buffer sizing.

14.3.4 Direct to device delivery

The previous section provides insight into how Urban Airship integrates with a third
party for message delivery. In referring to figure 14.1, note that two paths exist for
delivering messages to a device. In addition to delivering messages through a third
party, Urban Airship has experience serving directly as a channel for message deliv-
ery. In this capacity, individual devices connect directly to Urban Airship’s infrastruc-
ture, bypassing third-party providers. This approach brings a distinctly different set
of challenges:
Socket connections from mobile devices are often short-lived. Mobile devices frequently
switch between different types of networks depending on various conditions. To
back-end providers of mobile services, devices constantly reconnect and experi-
ence short but frequent periods of connectivity.
Connectivity across platforms is irregular. From a network perspective, tablet devices
tend to behave differently than mobile phones, and mobile phones behave dif-
ferently than desktop computers.
Frequency of mobile phone updates to back-end providers is certain to increase. Mobile
phones are increasingly used for daily tasks, producing significant amounts of
general network traffic but also analytics data for back-end providers.
Battery and bandwidth can’t be ignored. Unlike a traditional desktop environment,
mobile phones tend to operate on limited data plans. Service providers must
honor the fact that end users have limited battery life and they use expensive,
limited bandwidth. Abuse of either will frequently result in the uninstallation of
an application, the worst possible outcome for a mobile developer.
All aspects of infrastructure will need to scale massively. As mobile device popularity
increases, more application installations result in more connections to a mobile
services infrastructure. Each of the previous elements in this list are further
complicated by the sheer scale and growth of mobile devices.

Licensed to Marocco Aurélien <[email protected]>

106 CHAPTER 14 Case studies, part 1

Over time, Urban Airship learned several critical lessons as connections from mobile
devices continued to grow:
The diversity of mobile carriers can have a dramatic effect on device connectiv-
ity.
Many carriers don’t allow TCP keep-alive functionality. Given that, many carriers
will aggressively cull idle TCP sessions.
UDP isn’t a viable channel for messaging to mobile devices because many carri-
ers disallow it.
The overhead of SSLv3 is an acute pain for short-lived connections.
Given the challenges of mobile growth and the lessons learned by Urban Airship,
Netty was a natural fit for implementing a mobile messaging platform for reasons
highlighted in the following sections.

14.3.5 Netty excels at managing large numbers of concurrent

connections
As mentioned in the previous section, Netty makes supporting asynchronous I/O on the
JVM trivial. Because Netty operates on the JVM, and because the JVM on Linux ulti-
mately uses the Linux epoll facility to manage interest in socket file descriptors, Netty
makes it possible to accommodate the rapid growth of mobile by allowing developers to
easily accept large numbers of open sockets—close to 1 million TCP connections per
single Linux process. At numbers of this scale, service providers can keep costs low,
allowing a large number of devices to connect to a single process on a physical server.1
In controlled testing and with configuration options optimized to use small
amounts of memory, a Netty-based service was able to accommodate slightly less than
1 million connections (approximately 998,000). In this case, the limit was fundamen-
tally the Linux kernel imposing a hard-coded limit of 1 million file handles per pro-
cess. Had the JVM itself not held a number of sockets and file descriptors for JAR files,
the server would likely have been capable of handling even more connections, all on a
4 GB heap. Leveraging this efficiency, Urban Airship has successfully sustained over 20
million persistent TCP socket connections to its infrastructure for message delivery, all
on a handful of servers.
It’s worth noting that while in practice a single Netty-based service is capable of
handling nearly a million inbound TCP socket connections, doing so is not necessarily
pragmatic or advisable. As with all things in distributed computing, hosts will fail, pro-
cesses will need to be restarted, and unexpected behavior will occur. As a result of
these realities, proper capacity planning means considering the consequences of a sin-
gle process failing.

1
Note the distinction of a physical server in this case. Although virtualization offers many benefits, leading cloud
providers were regularly unable to accommodate more than 200,000–300,000 concurrent TCP connections
to a single virtual host. With connections at or above this scale, expect to use bare metal servers and expect to
pay close attention to the NIC (Network Interface Card) vendor.

Licensed to Marocco Aurélien <[email protected]>

 Summary 107

14.3.6 Summary—Beyond the perimeter of the firewall

We’ve demonstrated two everyday uses of Netty at the perimeter of the Urban Airship
network. Netty works exceptionally well for these purposes, but it has also found a
home as scaffolding for many other components inside Urban Airship.
INTERNAL RPC FRAMEWORK
Netty has been the centerpiece of an internal RPC framework that has consistently
evolved inside Urban Airship. Today, this framework processes hundreds of thousands
of requests per second with very low latency and exceptional throughput. Nearly every
API request fielded by Urban Airship processes through multiple back-end services
with Netty at the core of all of those services.
LOAD AND PERFORMANCE TESTING
Netty has been used at Urban Airship for several different load- and performance-test-
ing frameworks. For example, to simulate millions of device connections in testing the
previously described device-messaging service, Netty was used in conjunction with a
Redis (http://redis.io/) instance to test end-to-end message throughput with a mini-
mal client-side footprint.
ASYNCHRONOUS CLIENTS FOR COMMONLY SYNCHRONOUS PROTOCOLS
For some internal uses, Urban Airship has been experimenting with Netty to create
asynchronous clients for typically synchronous protocols, including services like
Apache Kafka (http://kafka.apache.org/) and Memcached (http://mem-
cached.org/). Netty’s flexibility easily allows us to craft clients that are asynchronous
in nature but that can be converted back and forth between truly asynchronous or syn-
chronous implementations without requiring upstream code changes.
All in all, Netty has been a cornerstone of Urban Airship as a service. The authors
and community are fantastic and have produced a truly first-class framework for any-
thing requiring networking on the JVM.

14.4 Summary
This chapter aimed at providing insight into real-world use of Netty and how it has
helped companies to solve significant networking problems. It’s worth noting how in
all cases Netty was leveraged not only as a code framework, but also as an essential
component of development and architectural best practices.
In the next chapter we’ll present case studies contributed by Facebook and Twitter
describing open source projects that evolved from Netty-based code originally devel-
oped to address internal needs.

Licensed to Marocco Aurélien <[email protected]>

108 CHAPTER 14 Case studies, part 1

Netty is a Java-based networking framework that man-

ages complex networking, multithreading, and concur-
rency for your applications. And Netty hides the
boilerplate and low-level code, keeping your business
logic separate and easier to reuse. With Netty, you get
an easy-to-use API, leaving you free to focus on what’s
unique to your application.
Netty in Action introduces the Netty framework and
shows you how to incorporate it into your Java network
applications. You will discover how to write highly scal-
able applications without getting into low-level APIs.
The book teaches you to think in an asynchronous way
as you work through its many hands-on examples and helps you master the best prac-
tices of building large-scale network apps.

What's inside
Netty from the ground up
Asynchronous, event-driven programming
Implementing services using different protocols
Covers Netty 4.x
This book assumes readers are comfortable with Java and basic network architecture.

Licensed to Marocco Aurélien <[email protected]>

 index
A AWS (Amazon Web Services)
account creation
Aandahaar, biometric database system 73 choosing support plan 26
AbstractRequestHandler 91 contact information 23
account, AWS creating key pair 28–31
creating login credentials 22–23
choosing support plan 26 payment details 24
contact information 23 signing in 26–28
creating key pair 28–31 verifying identity 24–25
login credentials 22–23 advantages of
payment details 24 automation capabilities 10
signing in 26–28 cost 11
verifying identity 24–25 fast-growing platform 9
AdMob mobile advertising network 73 platform of services 10
Advanced Messaging Queuing Model 69 reducing time to market 11
agents. See slaves reliability 11
Agoura Games, RabbitMQ and 73 scalability 10–11
allocation module 55, 63 standards compliance 11–12
Amazon Web Services. See AWS worldwide deployments 11
AMQ model See Advanced Message Queuing alternatives to 14–16
Model as cloud computing platform 3–4
AMQP specification 72 costs
platform and vendor neutral 70 billing example 12–13
AMQP, RabbitMQ and 71–72 Free Tier 12
Apache Kafka 107 overview 12
Apache ZooKeeper. See ZooKeeper pay-per-use pricing model 14
APNS (Apple Push Notification Service) 102 services overview 16–18
application tools for
loosely coupled 77 blueprints 21–22
routing-key attribute 81 CLI 19–21
scale-out approach 78 Management Console 19
tightly coupled 76 SDKs 21
application deployment 62–63 uses for
architectures, loosely coupled 74–76 data archiving 6–7
attributes 64 fault-tolerant systems 8

109

Licensed to Marocco Aurélien <[email protected]>

110 INDEX

running Java EE applications 5–6 deployment

running web shop 4–5 worldwide support 11
Azure 15–16 distributed architecture 63–65
frameworks 65
B masters 63
slaves 64–65
binding 80–82 distributed out of the box technologies 60
blueprints Docker
overview 21–22 images and containers 41
building Docker image 45–46 key commands 41
“by hand” method 43–44 overview 37–38
resources for 59
uses of 39–40
C
documenting software dependencies and
touchpoints 40
calculator for monthly costs 12
enabling continuous delivery 40
CDN (content delivery network) 5
enabling full-stack productivity when
cgroups. See control groups
offline 39
CLI (command-line interface)
enabling microservices architecture 39
overview 19–21
modelling networks 39
cloud computing
packaging software 39
overview 3–4
prototyping software 39
clusters 61–62 reducing debugging overhead 39
CMD command 45 replacing virtual machines 39
compute services 17 Docker application, building 42–50
containerizer 57 building Docker image 45–46
containers creating Docker image 43–44
overview 41, 58–59 layering 48–50
content delivery network. See CDN running Docker container 46–48
continuous delivery. See CD writing Dockerfile 44–45
control groups (cgroups) 54, 58–59, 62 docker build command 41
cost docker commit command 41, 43
advantages of AWS 11 Docker container 46–48
billing example 12–13 docker diff subcommand 48
Free Tier 12 docker run subcommand 48
overview 12 docker tag command 41
pay-per-use pricing model 14 Dockerfile method 43
cross-datacenter distribution of data 77 Dockerfiles
cross-node communication 72 writing 44–45
DRF (Dominant Resource Fairness) 56
D Droplr case study
creating faster uploads 86
data archiving 6–7 overview 85, 93
data centers performance 92
hardware used 3 technology stack
locations of 3, 11 handler implementations 91
data security standard. See DSS overview 88–89
databases parent interface 91
defined 18 request handlers 91
decoupling 74–75 server and pipeline 89–91
deploying applications upload request handler 92
overview 62–63 DSS (data security standard) 12

Licensed to Marocco Aurélien <[email protected]>

 INDEX 111

E K
EC2 (Elastic Compute Cloud) service kernel 63
defined 2 key pair for SSH
See also virtual servers creating 28–31
Elastic Compute Cloud service. See EC2 service
enterprise services 18 L
Ericsson Computer Science Laboratory 71
Erlang, RabbitMQ and 71 layering 41, 48–50
exchange, component of the message broker 80 leading master 63
type of 81 Linux
executors 64–65
key file permissions 30
Exited status 48
EXPOSE command 45
M
F Mac OS X
key file permissions 30
fault-tolerance
AWS use cases 8 MAINTAINER command 44
Firebase case study Management Console
HTTP 1.1 keep-alive 97–98 overview 19
long polling 95–97 signing in 26
overview 93–94, 100 masters
SSL handler control 98–100 overview 56, 63
first-in, first-out (FIFO) order 81 Memcached 107
frameworks 63, 65 MercadoLibre, e-commerce ecosystem 73
Free Tier 12 Mesos framework
FROM command 44 distributred architecture of 63–65
frameworks 65
G masters 63
slave 64
git 45 slaves 64–65
Google Cloud Platform 15–16 how it works 55–57
resource isolation 57
H resource offers 56
two-tier scheduling 56–57
hardware 3 when to use 59–60
HTTP 1.1 keep-alive 97–98 message broker 70–71
tradeoffs and introducing to an architecture 78
I messaging-based architectures, RabbitMQ and 69
messaging-oriented-middleware (MOM)
IaaS (infrastracture as a service) 4 defined 75
images operational complexity and 78
building 45–46 RabbitMQ as 69
overview 41 multitenancy 54, 57
infrastracture as a service. See IaaS
Internet, RabbitMQ and 70 N
inter-process communication (IPC) system 71
NASA 73
J NAT (Network Address Translation) 5
NioServerSocketChannelFactory 89
Java EE applications 5–6 Node.js image 44

Licensed to Marocco Aurélien <[email protected]>

112 INDEX

O resource partioning 61–62

resources 64
Ocean Observations Initiative 73 RUN command 45
OpenStack 14–16 running Docker container 46–48
operating system and environments, RabbitMQ
and sharing data across 70 S
P S3 (Simple Storage Service)
defined 2
PaaS (platform as a service) 4 SaaS (software as a service) 4
partitioning resources. See resource partioning scaling
PCI (payment card industry) 12 advantages of AWS 10–11
performance degradation 75 SDKs (software development kits)
database updates and 75 overview 21
performance, Droplr case study 92 security, RabbitMQ and 70
Pivotal Software Inc. 70 SimpleHandler 92
platform as a service. See PaaS slaves
plugin extensibility, RabbitMQ and 72 overview 56, 64
protocol neutrality, RabbitMQ and 72 software as a service. See SaaS
ps command 48 software development kits. See SDKs
PuTTY 30–31 SslHandler class 98
Python 82 standards compliance 11–12
stateful technologies 60
Q-R stateless technologies 60
storage
queue 80–81 defined 18
RabbitMQ
advanced routing and message distribution 75
AMQP specification and 80
T
and adding new functionality seamlessly 77
third-party plugins 70
and bi-direction federation of data 78
timestamp 74–75
and multiple exchange types 80
to-do application 42, 50
application decoupling 76
tools
as an open-source project 70
blueprints 21–22
binding keys and 81
CLI 19–21
clusters See clusters
Management Console 19
data and event replication 77–79
database write decoupling 77 SDKs 21
described 69 two-tier scheduling 56
diverse functionality 82
loosely coupled data See also decoupling U
loosely coupled design 76
the most fundamental features of 71 Urban Airship case study
unique features 69–71 binary protocol example 102–105
RabbitMQ federation plugin 78 direct to device delivery 105–106
rabbitpy, library 82 large numbers of concurrent connections 106
Rapportive, GMail add-on 73 overview 100–102, 107
RDP (Remote Desktop Protocol) 28 third-party delivery 102
real-time system 71 use cases
Reddit, online community 73 data archiving 6–7
reliability 11 fault-tolerant systems 8
Remote Desktop Protocol. See RDP running Java EE applications 5–6
resource offers 63 running web shop 4–5

Licensed to Marocco Aurélien <[email protected]>

 INDEX 113

V W-Z
virtual machines, comparing containers to 57–58 WebSocket
virtual machines. See VMs defined 98
Virtual Memory (VM) 72 Droplr case study 95
VMs (virtual machines) 6 Windows
VPN (Virtual Private Network) 5 SSH client on 30–31
WORKDIR command 45
writing, Dockerfile 44–45
ZooKeeper 63

Licensed to Marocco Aurélien <[email protected]>

Save 50% on these selected books—eBook, pBook, and MEAP. Just enter fenative in the Pro-
motional Code box when you check out. Only at manning.com.

Amazon Web Services in Action

by Michael Wittig and Andreas Wittig
ISBN: 9781617292880
424 pages
$49.99
September 2015

Docker in Practice
by Ian Miell and Aidan Hobson Sayers
ISBN: 9781617292729
325 pages
$44.99
March 2016

Mesos in Action
by Roger Ignazio
ISBN: 9781617292927
325 pages
$44.99
April 2016

Licensed to Marocco Aurélien <[email protected]>

RabbitMQ in Depth
by Gavin M. Roy
ISBN: 9781617291005
375 pages
$59.99
May 2016

Netty in Action
by Norman Maurer and Marvin Allen Wolfthal
ISBN: 9781617291470
296 pages
$54.99
September 2015