Professional Documents
Culture Documents
Data Engineering Cookbook
Data Engineering Cookbook
Andreas Kretz
July 2, 2019
v2.1
Contents
I Introduction 10
3 Learn To Code 17
2
7 Computer Networking - Data Transmission 25
7.1 OSI Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
7.2 IP Subnetting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
7.3 Switch, Level 3 Switch . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
7.4 Router . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
7.5 Firewalls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
9 Linux 28
9.1 OS Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
9.2 Shell scripting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
9.3 Cron jobs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
9.4 Packet management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
10 The Cloud 30
10.1 IaaS vs PaaS vs SaaS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
10.2 AWS,Azure, IBM, Google Cloud basics . . . . . . . . . . . . . . . . . . . 30
10.3 Cloud vs On-Premises . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
10.4 Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
10.5 Hybrid Clouds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
12 Big Data 33
12.1 What is big data and where is the difference to data science and data
analytics? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
12.2 The 4Vs of Big Data — available . . . . . . . . . . . . . . . . . . . . . . 33
12.3 Why Big Data? — available . . . . . . . . . . . . . . . . . . . . . . . . . 34
12.3.1 Planning is Everything . . . . . . . . . . . . . . . . . . . . . . . . 35
12.3.2 The Problem With ETL . . . . . . . . . . . . . . . . . . . . . . . 35
12.3.3 Scaling Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3
12.3.4 Scaling Out . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
12.3.5 Please Don’t go Big Data . . . . . . . . . . . . . . . . . . . . . . 38
14 Lambda Architecture 43
14.1 Batch Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
14.2 Stream Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
14.3 Should you do stream or batch processing? . . . . . . . . . . . . . . . . . 45
14.4 Lambda Architecture Alternative . . . . . . . . . . . . . . . . . . . . . . 45
14.4.1 Kappa Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 45
14.4.2 Kappa Architecture with Kudu . . . . . . . . . . . . . . . . . . . 45
14.5 Why a Good Data Platform Is Important . . . . . . . . . . . . . . . . . . 45
17 Docker 52
17.1 What is docker and what do you use it for — available . . . . . . . . . . 52
17.1.1 Don’t Mess Up Your System . . . . . . . . . . . . . . . . . . . . . 52
17.1.2 Preconfigured Images . . . . . . . . . . . . . . . . . . . . . . . . . 52
17.1.3 Take It With You . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
17.2 Kubernetes Container Deployment . . . . . . . . . . . . . . . . . . . . . 53
17.3 How to create, start,stop a Container . . . . . . . . . . . . . . . . . . . . 54
17.4 Docker micro services? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
17.5 Kubernetes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
17.6 Why and how to do Docker container orchestration . . . . . . . . . . . . 54
4
17.7 Useful Docker Commands . . . . . . . . . . . . . . . . . . . . . . . . . . 54
18 REST APIs 56
18.1 API Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
18.2 Implementation Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . 56
18.3 OAuth security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
19 Databases 58
19.1 SQL Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
19.1.1 PostgreSQL DB . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
19.1.2 Database Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
19.1.3 SQL Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
19.1.4 Stored Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . 58
19.1.5 ODBC/JDBC Server Connections . . . . . . . . . . . . . . . . . . 58
19.2 NoSQL Stores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
19.2.1 KeyValue Stores (HBase) . . . . . . . . . . . . . . . . . . . . . . . 58
19.2.2 Document Store HDFS — available . . . . . . . . . . . . . . . . . 58
19.2.3 Document Store MongoDB . . . . . . . . . . . . . . . . . . . . . . 61
19.2.4 Elasticsearch Search Engine and Document Store . . . . . . . . . 62
19.2.5 Hive Warehouse . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
19.2.6 Impala . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
19.2.7 Kudu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
19.2.8 Apache Druid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
19.2.9 InfluxDB Time Series Database . . . . . . . . . . . . . . . . . . . 63
19.2.10 MPP Databases (Greenplum) . . . . . . . . . . . . . . . . . . . . 64
5
20.4.2 How does Spark fit to Hadoop? – available . . . . . . . . . . . . . 72
20.4.3 Where’s the difference? . . . . . . . . . . . . . . . . . . . . . . . . 72
20.4.4 Spark and Hadoop is a perfect fit . . . . . . . . . . . . . . . . . . 73
20.4.5 Spark on YARN: . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
20.4.6 My simple rule of thumb: . . . . . . . . . . . . . . . . . . . . . . 74
20.4.7 Available Languages – available . . . . . . . . . . . . . . . . . . . 74
20.4.8 How to do stream processing . . . . . . . . . . . . . . . . . . . . . 74
20.4.9 How to do batch processing . . . . . . . . . . . . . . . . . . . . . 74
20.4.10 How does Spark use data from Hadoop – available . . . . . . . . . 74
20.4.11 What is a RDD and what is a DataFrame? . . . . . . . . . . . . . 75
20.4.12 Spark coding with Scala . . . . . . . . . . . . . . . . . . . . . . . 75
20.4.13 Spark coding with Python . . . . . . . . . . . . . . . . . . . . . . 75
20.4.14 How and why to use SparkSQL? . . . . . . . . . . . . . . . . . . . 75
20.4.15 Machine Learning on Spark? (Tensor Flow) . . . . . . . . . . . . 75
20.4.16 MLlib: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
20.4.17 Spark Setup – available . . . . . . . . . . . . . . . . . . . . . . . . 76
20.4.18 Spark Resource Management – available . . . . . . . . . . . . . . 76
20.5 Apache Nifi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
20.6 StreamSets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
21 Apache Kafka 78
21.1 Why a message queue tool? . . . . . . . . . . . . . . . . . . . . . . . . . 78
21.2 Kakfa architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
21.3 What are topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
21.4 What does Zookeeper have to do with Kafka . . . . . . . . . . . . . . . . 78
21.5 How to produce and consume messages . . . . . . . . . . . . . . . . . . . 78
21.6 KAFKA Commands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
22 Machine Learning 80
22.1 Training and Applying models . . . . . . . . . . . . . . . . . . . . . . . . 80
22.2 What is deep learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
22.3 How to do Machine Learning in production — available . . . . . . . . . . 80
22.4 Why machine learning in production is harder then you think – available 81
22.5 Models Do Not Work Forever . . . . . . . . . . . . . . . . . . . . . . . . 81
22.6 Where The Platforms That Support This? . . . . . . . . . . . . . . . . . 81
22.7 Training Parameter Management . . . . . . . . . . . . . . . . . . . . . . 82
22.8 What’s Your Solution? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
22.9 How to convince people machine learning works — available . . . . . . . 82
22.10No Rules, No Physical Models . . . . . . . . . . . . . . . . . . . . . . . . 83
6
22.11You Have The Data. USE IT! . . . . . . . . . . . . . . . . . . . . . . . . 83
22.12Data is Stronger Than Opinions . . . . . . . . . . . . . . . . . . . . . . . 84
22.13AWS Sagemaker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
23 Data Visualization 85
23.1 Android & IOS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
23.2 How to design APIs for mobile apps . . . . . . . . . . . . . . . . . . . . . 85
23.3 How to use Webservers to display content . . . . . . . . . . . . . . . . . 85
23.3.1 Tomcat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
23.3.2 Jetty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
23.3.3 NodeRED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
23.3.4 React . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
23.4 Business Intelligence Tools . . . . . . . . . . . . . . . . . . . . . . . . . . 86
23.4.1 Tableau . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
23.4.2 PowerBI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
23.4.3 Quliksense . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
23.5 Identity & Device Management . . . . . . . . . . . . . . . . . . . . . . . 86
23.5.1 What is a digital twin? . . . . . . . . . . . . . . . . . . . . . . . . 86
23.5.2 Active Directory . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
24 What We Want To Do 88
29 Apache Zeppelin 93
29.1 Install and Ingest Kafka Topic . . . . . . . . . . . . . . . . . . . . . . . . 93
29.2 Processing Messages with Spark & SparkSQL . . . . . . . . . . . . . . . 93
29.3 Visualizing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7
30.4 Move Zeppelin Code to Spark . . . . . . . . . . . . . . . . . . . . . . . . 94
IV Case Studies 95
8
31.33Data Science @Upwork . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
31.34Data Science @Woot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
31.35Data Science @Zalando . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
9
Part I
Introduction
10
1 How To Use This Cookbook
What do you actually need to learn to become an awesome data engineer? Look no
further, you’ll find it here.
If you are looking for AI algorithms and such data scientist things, this book is not for
you.
You are going to find Five Types of Content in this book: Articles I wrote, links to
my podcast episodes (video & audio), more then 200 links to helpful websites I like, data
engineering interview questions and case studies.
Check out this podcast episode where I talk in detail why I decided to share all this
information for free: #079 Trying to stay true to myself and making the cookbook public
on GitHub
11
2 Data Engineer vs Data Scientists
Podcast Episode: #050 Data Engineer Scientist or Analyst Which One Is For You?
In this podcast we talk about the differences between data scientists, analysts and
engineers. Which are the three main data science jobs. All three super important.
This makes it easy to decide
YouTube Click here to watch
Audio Click here to listen
Table 2.1: Podcast: 050 Data Engineer Scientist or Analyst Which One Is For You?
Data scientists do not wear white coats or work in high tech labs full of science fiction
movie equipment. They work in offices just like you and me.
What differs them from most of us is that they are the math experts. They use linear
algebra and multivariable calculus to create new insight from existing data.
Here’s an example:
An industrial company produces a lot of products that need to be tested before shipping.
Usually such tests take a lot of time because there are hundreds of things to be tested.
All to make sure that your product is not broken.
Wouldn’t it be great to know early if a test fails ten steps down the line? If you knew
that you could skip the other tests and just trash the product or repair it.
That’s exactly where a data scientist can help you, big-time. This field is called predictive
analytics and the technique of choice is machine learning.
12
Yes, machine learning, it works like this:
You feed an algorithm with measurement data. It generates a model and optimises it
based on the data you fed it with. That model basically represents a pattern of how your
data is looking You show that model new data and the model will tell you if the data
still represents the data you have trained it with. This technique can also be used for
predicting machine failure in advance with machine learning. Of course the whole process
is not that simple.
The actual process of training and applying a model is not that hard. A lot of work
for the data scientist is to figure out how to pre-process the data that gets fed to the
algorithms.
Because to train a algorithm you need useful data. If you use any data for the training
the produced model will be very unreliable.
A unreliable model for predicting machine failure would tell you that your machine is
damaged even if it is not. Or even worse: It would tell you the machine is ok even when
there is an malfunction.
Model outputs are very abstract. You also need to post-process the model outputs to
receive health values from 0 to 100.
Data Engineers are the link between the management’s big data strategy and the data
scientists that need to work with data.
What they do is building the platforms that enable data scientists to do their magic.
13
• Algorithm creation by data scientists
• Automation of the data scientist’s machine learning models and algorithms for
production use
• Most of the time these guys start as traditional solution architects for systems
that involve SQL databases, web servers, SAP installations and other “standard”
systems.
But to create big data platforms the engineer needs to be an expert in specifying, set-
ting up and maintaining big data technologies like: Hadoop, Spark, HBase, Cassandra,
MongoDB, Kafka, Redis and more.
What they also need is experience on how to deploy systems on cloud infrastructure like
at Amazon or Google or on premise hardware.
Table 2.2: Podcast: 048 From Wannabe Data Scientist To Engineer My Journey
For a good company it is absolutely important to get well trained data engineers and data
scientists. Think of the data scientist as the professional race car driver. A fit athlete
with talent and driving skills like you have never seen.
What he needs to win races is someone who will provide him the perfect race car to drive.
That’s what the solution architect is for.
Like the driver and his team the data scientist and the data engineer need to work closely
together. They need to know the different big data tools Inside and out.
That’s why companies are looking for people with Spark experience. It is a common
ground between both that drives innovation.
Spark gives data scientists the tools to do analytics and helps engineers to bring the data
scientist’s algorithms into production. After all, those two decide how good the data
14
platform is, how good the analytics insight is and how fast the whole system gets into a
production ready state.
15
Part II
16
3 Learn To Code
Why this is important: Without coding you cannot do much in data engineering. I cannot
count the number of times I needed a quick Java hack.
When you are getting into data processing with Spark you should use Scala. But, after
learning Java this is easy to do.
Personally however, I am not that big into Python. But I am going to look into it
Where to Learn? There’s a Java Course on Udemy you could look at: https://www.udemy.com/java-
programming-tutorial-for-beginners
• What are Unit tests to make sure what you code is working
• Functional Programming
17
4 Get Familiar With Git
Why this is important: One of the major problems with coding is to keep track of changes.
It is also almost impossible to maintain a program you have multiple versions of.
Let’s say you work on a Spark application and your colleagues need to make changes
while you are on holiday. Without some code management they are in huge trouble:
Where is the code? What have you changed last? Where is the documentation? How do
we mark what we have changed?
But if you put your code on GitHub your colleagues can find your code. They can
understand it through your documentation (please also have in-line comments)
Developers can pull your code, make a new branch and do the changes. After your holiday
you can inspect what they have done and merge it with your original code. and you end
up having only one application
Where to learn: Check out the GitHub Guides page where you can learn all the basics:
https://guides.github.com/introduction/flow/
This great GitHub commands cheat sheet saved my butt multiple times: https://www.atlassian.com/git/
git-cheatsheet
• Pull
• Push
• Branching
• Forking
18
5 Agile Development – available
These days everyone wants to be agile. Big or small company people are looking for the
“startup mentality”.
Many think it’s the corporate culture. Others think it’s the process how we create things
that matters.
In this article I am going to talk about agility and self-reliance. About how you can
incorporate agility in your professional career.
It’s a bit of an arrogant process. You assume that you already know exactly what a
customer wants. Or how a product has to look and how everything works out.
The problem is that the world does not work this way!
Sometimes things just do not work out as planned or stuff is harder than you think.
Other times you find out that you build something customers do not like and need to be
changed.
That’s why people jump on the Scrum train. Because Scrum is the definition of agile
development, right?
19
5.2 Agile rules I learned over the years – available
Yes, Scrum or Google’s OKR can help to be more agile. The secret to being agile however,
is not only how you create.
What makes me cringe is people try to tell you that being agile starts in your head. So,
the problem is you?
No!
The biggest lesson I have learned over the past years is this: Agility goes down the drain
when you outsource work.
I know on paper outsourcing seems like a no brainer: Development costs against the fixed
costs.
It is expensive to bind existing resources on a task. It is even more expensive if you need
to hire new employees.
The problem with outsourcing is that you pay someone to build stuff for you.
It does not matter who you pay to do something for you. He needs to make money.
His agenda will be to spend as less time as possible on your work. That is why outsourcing
requires contracts, detailed specifications, timetables and delivery dates.
He doesn’t want to spend additional time on a project, only because you want changes
in the middle. Every unplanned change costs him time and therefore money.
If so, you need to make another detailed specification and a contract change.
He is not going to put his mind into improving the product while developing. Firstly
because he does not have the big picture. Secondly because he does not want to.
He is doing as he is told.
Who can blame him? If I was the subcontractor I would do exactly the same!
20
5.2.3 Knowledge is king: A lesson from Elon Musk
Doing everything in house, that’s why startups are so productive. No time is wasted on
waiting for someone else.
If something does not work, or needs to be changed, there is someone in the team who
can do it right away. .
One very prominent example who follows this strategy is Elon Musk.
Tesla’s Gigafactories are designed to get raw materials in on one side and spit out cars
on the other. Why do you think Tesla is building Gigafactories who cost a lot of money?
Why is SpaceX building its one space engines? Clearly there are other, older, companies
who could do that for them.
Why is Elon building tunnel boring machines at his new boring company?
If you look closer it all comes down to control and knowledge. You, your team, your
company, needs to do as much as possible on your own. Self-reliance is king.
Build up your knowledge and therefore the teams knowledge. When you have the ability
to do everything yourself, you are in full control.
Don’t largely rely on others and be confident to just do stuff on your own.
PS. Don’t get me wrong. You can still outsource work. Just do it in a smart way by
outsourcing small independent parts.
21
5.3 Agile Frameworks
5.3.1 Scrum
There’s a interesting Scrum Medium publication with a lot of details about Scrum: https:
//medium.com/serious-scrum
Also this scrum guide webpage has good infos about Scrum: https://www.scrumguides.
org/scrum-guide.html
5.3.2 OKR
I personally love OKR, been doing it for years. Especially for smaller teams OKR is
great. You don’t have a lot of overhead and get work done. It helps you stay focused
and look at the bigger picture.
I recommend to do a sync meeting every Monday. There you talk about what happened
last week and what you are going to work on this week.
This is also this awesome 1,5 hours Startup guide from Google: https://youtu.be/mJB83EZtAjc
I really love this video, I rewatched it multiple times.
The software engineering and development culture is super important. How does a com-
pany handle product development with hundreds of developers. Check out this podcast:
https://labs.spotify.com/2014/03/27/spotify-engineering-culture-part-1/
22
https://labs.spotify.com/2014/09/20/spotify-engineering-culture-part-2/
23
6 Learn how a Computer Works
6.1 CPU,RAM,GPU,HDD
I talked about computer hardware and GPU processing in this podcast: https://anchor.
fm/andreaskayy/embed/episodes/Why-the-hardware-and-the-GPU-is-super-important--PoDS-030-e23
24
7 Computer Networking - Data
Transmission
The OSI Model describes how data is flowing through the network. It consists of layers
starting from physical layers, basically how the data is transmitted over the line or optic
fiber.
Cisco page that shows the layers of the OSI model and how it works: https://learningnetwork.
cisco.com/docs/DOC-30382
Which protocol lives on which layer? Check out this network protocol map. Un-
fortunately it is really hard to find it theses days: https://www.blackmagicboxes.com/
wp-content/uploads/2016/12/Network-Protocols-Map-Poster.jpg
7.2 IP Subnetting
Check out this IP Adress and Subnet guide from Cisco: https://www.cisco.com/c/en/
us/support/docs/ip/routing-information-protocol-rip/13788-3.html
25
7.3 Switch, Level 3 Switch
7.4 Router
7.5 Firewalls
26
8 Security and Privacy
27
9 Linux
Linux is very important to learn, at least the basics. Most Big Data tools or NoSQL
databases are running on Linux.
From time to time you need to modify stuff through the operation system. Especially
if you run a infrastructure as a service solution like Cloudera CDH, Hortonworks or a
MapR Hadoop distribution
9.1 OS Basics
Cron jobs are super important to automate simple processes or jobs in Linux. You need
this here and there I promise. Check out this three guides:
https://linuxconfig.org/linux-crontab-reference-guide
https://www.ostechnix.com/a-beginners-guide-to-cron-jobs/
Pro tip: Don’t forget to end your cron files with an empty line or a comment, otherwise
it will not work.
28
9.4 Packet management
29
10 The Cloud
Check out this Podcast it will help you understand where’s the difference and how to
decide on what you are going to use.
Podcast Episode: #082 Reading Tweets With Apache Nifi & IaaS vs PaaS vs SaaS
In this episode we are talking about the differences between infrastructure as a
service, platform as a service and application as a service. Then we install the Nifi
docker container and look into how we can extract the twitter data.
Youtube Click here to watch
Audio Click here to listen
Table 10.1: Podcast: 082 Reading Tweets With Apache Nifi & IaaS vs PaaS vs SaaS
30
10.4 Security
Hybrid clouds are a mixture of on-premises and cloud deployment. A very interesting
example for this is Google Anthos:
https://cloud.google.com/anthos/
31
11 Security Zone Design
I talked about security zone design and lambda architecture in this podcast: https://
anchor.fm/andreaskayy/embed/episodes/How-to-Design-Security-Zones-and-Lambda-Architecture--Po
32
12 Big Data
It is a complete misconception. Volume is only one part of the often called four V’s of
big data: Volume, velocity, variety and veracity.
Velocity is about the speed of how fast the data is getting to you.
How much data is in a specific time needs to get processed or is coming into the system.
This is where the whole concept of streaming data and real-time processing comes in to
play.
Variety is the third one. It means, that the data is very different. That you have very
different types of data structures.
Like CSV files, PDFs that you have stuff in XML. That you have JSON logfiles, or that
you have data in some kind of a key value store.
It’s about the variety of data types from different sources that you basically want to join
together. All to make an analysis based on that data.
Veracity is fourth and this is a very very difficult one. The issue with big data is, that
it is very unreliable.
You cannot really trust the data. Especially when you’re coming from the IoT, the
33
Internet of Things side. Devices use sensors for measurement of temperature, pressure,
acceleration and so on.
You cannot always be hundred percent sure that the actual measurement is right.
When you have data that is from for instance SAP and it contains data that is created
by hand you also have problems. As you know we humans are bad at inputting stuff.
Everybody articulates different. We make mistakes, down to the spelling and that can
be a very difficult issue for analytics.
What I always emphasize is the four V’s are quite nice. They give you a general direction.
What I mean by catastrophic success is, that your project, your startup or your platform
has more growth that you anticipated. Exponential growth is what everybody is looking
for.
Because with exponential growth there is the money. It starts small and gets very big
very fast. The classic hockey stick curve:
1,2,4,8,16,32,64,128,256,512,1024,2048,4096,8192,16384. . . .BOOM!
Think about it. It starts small and quite slow, but gets very big very fast.
You get a lot of users or customers who are paying money to use your service, the platform
or whatever. If you have a system that is not equipped to scale and process the data the
whole system breaks down.
That’s catastrophic success. You are so successful and grow so fast that you cannot fulfill
the demand anymore. And so you fail and it’s all over.
It’s now like you just can make that up while you go. That you can foresee in a few
months or weeks the current system doesn’t work anymore.
34
12.3.1 Planning is Everything
It’s all happens very very fast and you cannot react anymore. There’s a necessary type
of planning and analyzing the potential of your business case necessary.
Then you need to decide if you actually have big data or not.
You need to decide if you use big data tools. This means when you conceptualize the
whole infrastructure it might look ridiculous to actually focus on big data tools.
But in the long run it will help you a lot. Good planning will get a lot of problems out
of the way, especially if you think about streaming data and real-time analytics.
A typical old-school platform deployment would look like the picture below. Devices use
a data API to upload data that gets stored in a SQL database. An external analytics
tool is querying data and uploading the results back to the SQL db. Users then use the
user interface to display data stored in the database.
Now, when the front end queries data from the SQL database the following three steps
happen:
35
The database extracts all the needed rows from the storage. Extracted data gets trans-
formed, for instance sorted by timestamp or something a lot more complex.
The extracted and transformed data gets loaded to the destination (the user interface)
for chart creation With exploding amounts of stored data the ETL process starts being
a real problem.
Analytics is working with large data sets, for instance whole days, weeks, months or more.
Data sets are very big like 100GB or Terabytes. That means Billions or Trillions of rows.
This has the result that the ETL process for large data sets takes longer and longer. Very
quickly the ETL performance gets so bad it won’t deliver results to analytics anymore.
12.3.3 Scaling Up
To scale up the system and therefore increase ETL speeds administrators resort to more
powerful hardware by:
Speeding up the extract performance by adding faster disks to physically read the data
faster. Increasing RAM for row caching. What is already in memory does not have to be
read by slow disk drives. Using more powerful CPU’s for better transform performance
(more RAM helps here as well) Increasing or optimising networking performance for faster
data delivery to the front end and analytics Scaling up the system is fairly easy.
But with exponential growth it is obvious that sooner or later (more sooner than later)
you will run into the same problems again. At some point you simply cannot scale up
anymore because you already have a monster system, or you cannot afford to buy more
expensive hardware.
36
12.3.4 Scaling Out
Scaling out is the opposite of scaling up. Instead of building bigger systems the goal is
to distribute the load between many smaller systems.
The simplest way of scaling out an SQL database is using a storage area network (SAN)
to store the data. You can then use up to eight SQL servers, attach them to the SAN
and let them handle queries. This way load gets distributed between those eight servers.
One major downside of this setup is that, because the storage is shared between the
sql servers, it can only be used as an read only database. Updates have to be done
periodically, for instance once a day. To do updates all SQL servers have to detach from
the database. Then, one is attaching the db in read-write mode and refreshing the data.
This procedure can take a while if a lot of data needs to be uploaded.
This Link to a Microsoft MSDN page has more options of scaling out an SQL database
for you.
I deliberately don’t want to get into details about possible scaling out solutions. The
point I am trying to make is that while it is possible to scale out SQL databases it is very
complicated.
There is no perfect solution. Every option has its up- and downsides. One common
major issue is the administrative effort that you need to take to implement and maintain
37
a scaled out solution.
If you don’t run into scaling issues please, do not use big data tools!
Big data is a expensive thing. A Hadoop cluster for instance needs at least five servers
to work properly. More is better.
Especially when you are talking the maintenance and development on top big data tools
into account.
On the other side: If you really need big data tools they will save your ass :)
38
13 My Big Data Platform Blueprint
Some time ago I have created a simple and modular big data platform blueprint for
myself. It is based on what I have seen in the field and read in tech blogs all over the
internet.
Following my blueprint will allow you to create the big data platform that fits exactly your
needs. Building the perfect platform will allow data scientists to discover new insights.
It will enable you to perfectly handle big data and allow you to make data driven decisions.
THE BLUEPRINT The blueprint is focused on the four key areas: Ingest, store,
analyse and display.
39
Having the platform split like this turns it it a modular platform with loosely coupled
interfaces.
If you have a platform that is not modular you end up with something that is fixed or
hard to modify. This means you can not adjust the platform to changing requirements
of the company.
Because of modularity it is possible to switch out every component, if you need it.
13.1 Ingest
Ingestion is all about getting the data in from the source and making it available to later
stages. Sources can be everything form tweets, server logs to IoT sensor data like from
cars.
Sources send data to your API Services. The API is going to push the data into a
temporary storage.
The temporary storage allows other stages simple and fast access to incoming data.
A great solution is to use messaging queue systems like Apache Kafka, RabbitMQ or
AWS Kinesis. Sometimes people also use caches for specialised applications like Redis.
A good practice is that the temporary storage follows the publish, subscribe pattern.
This way APIs can publish messages and Analytics can quickly consume them.
The analyse stage is where the actual analytics is done. Analytics, in the form of stream
and batch processing.
Streaming data is taken from ingest and fed into analytics. Streaming analyses the “live”
data thus, so generates fast results.
As the central and most important stage, analytics also has access to the big data storage.
Because of that connection, analytics can take a big chunk of data and analyse it.
40
This type of analysis is called batch processing. It will deliver you answers for the big
questions.
To learn more about stream and batch processing read my blog post: How to Create New
and Exciting Big Data Aided Products
The analytics process, batch or streaming, is not a one way process. Analytics also can
write data back to the big data storage.
Often times writing data back to the storage makes sense. It allows you to combine
previous analytics outputs with the raw data.
Analytics insight can give meaning to the raw data when you combine them. This
combination will often times allow you to create even more useful insight.
A wide variety of analytics tools are available. Ranging from MapReduce or AWS Elastic
MapReduce to Apache Spark and AWS lambda.
13.3 Store
This is the typical big data storage where you just store everything. It enables you to
analyse the big picture.
Most of the data might seem useless for now, but it is of upmost importance to keep it.
Throwing data away is a big no no.
Although it seems useless for now, data scientists can work with the data. They might
find new ways to analyse the data and generate valuable insight from it.
Systems like Hadoop HDFS, Hbase, Amazon S3 or DynamoDB are a perfect fit to store
big data.
Check out my podcast how to decide between SQL and NoSQL: https://anchor.fm/andreaskayy/embed/
Vs-SQL-How-To-Choose-e12f1o
41
13.4 Display
Displaying data is as important as ingesting, storing and analysing it. People need to be
able to make data driven decisions.
This is why it is important to have a good visual presentation of the data. Sometimes
you have a lot of different use cases or projects using the platform.
It might not be possible for you to build the perfect UI that fits everyone. What you
should do in this case is enable others to build the perfect UI themselves.
How to do that? By creating APIs to access the data and making them available to
developers.
Either way, UI or API the trick is to give the display stage direct access to the data in
the big data cluster. This kind of access will allow the developers to use analytics results
as well as raw data to build the the perfect application.
42
14 Lambda Architecture
Ask the big questions. Remember your last yearly tax statement?
You break out the folders. You run around the house searching for the receipts.
When you finally found everything you fill out the form and send it on its way.
Data comes in and gets stored, analytics loads the data from storage and creates an
output (insight):
43
To do so, batch processing jobs use large amounts of data. This data is provided by
storage systems like Hadoop HDFS.
Results from batch jobs are very useful, but the execution time is high. Because the
amount of used data is high.
It can take minutes or sometimes hours until you get your results.
Streaming allows users to make quick decisions and take actions based on “real-time”
insight. Contrary to batch processing, streaming processes data on the fly, as it comes
in.
With streaming you don’t have to wait minutes or hours to get results. You gain instant
insight into your data.
In the batch processing pipeline, the analytics was after the data storage. It had access
to all the available data.
Stream processing creates insight before the data storage. It has only access to fragments
of data as it comes in.
As a result the scope of the produced insight is also limited. Because the big picture is
missing.
Only with streaming analytics you are able to create advanced services for the customer.
Netflix for instance incorporated stream processing into Chuckwa V2.0 and the new
Keystone pipeline.
One example of advanced services through stream processing is the Netflix “Trending
Now” feature. Check out the Netflix case study.
44
14.3 Should you do stream or batch processing?
It is a good idea to start with batch processing. Batch processing is the foundation of
every good big data platform.
A batch processing architecture is simple, and therefore quick to set up. Platform sim-
plicity means, it will also be relatively cheap to run.
A batch processing platform will enable you to quickly ask the big questions. They will
give you invaluable insight into your data and customers.
When the time comes and you also need to do analytics on the fly, then add a streaming
pipeline to your batch processing big data platform.
Podcast Episode: #066 How To Do Data Science From A Data Engineers Perspective
A simple introduction how to do data science in the context of the internet of things.
YouTube Click here to watch
Audio Click here to listen
Table 14.2: Podcast: 066 How To Do Data Science From A Data Engineers..
45
15 Data Warehouse vs Data Lake
46
16 Hadoop Platforms — available
When people talk about big data, one of the first things come to mind is Hadoop. Google’s
search for Hadoop returns about 28 million results.
It seems like you need Hadoop to do big data. Today I am going to shed light onto why
Hadoop is so trendy.
You will see that Hadoop has evolved from a platform into an ecosystem. It’s design
allows a lot of Apache projects and 3rd party tools to benefit from Hadoop.
I will conclude with my opinion on, if you need to learn Hadoop and if Hadoop is the
right technology for everybody.
Hadoop is a platform for distributed storing and analyzing of very large data sets.
Hadoop has four main modules: Hadoop common, HDFS, MapReduce and YARN. The
way these modules are woven together is what makes Hadoop so successful.
The Hadoop common libraries and functions are working in the background. That’s why
I will not go further into them. They are mainly there to support Hadoop’s modules.
Podcast Episode: #060 What Is Hadoop And Is Hadoop Still Relevant In 2019?
A Introduction into Hadoop HDFS, YARN and MapReduce. Yes, Hadoop is still
relevant in 2019 even if you look into serverless tools.
YouTube Click here to watch
Audio Click here to listen
Table 16.1: Podcast: 060 What Is Hadoop And Is Hadoop Still Relevant In 2019?
47
16.2 What makes Hadoop so popular? — available
Storing and analyzing data as large as you want is nice. But what makes Hadoop so
popular?
Hadoop’s core functionality is the driver of Hadoop’s adoption. Many Apache side
projects use it’s core functions.
Because of all those side projects Hadoop has turned more into an ecosystem. An ecosys-
tem for storing and processing big data.
To better visualize this eco system I have drawn you the following graphic. It shows some
projects of the Hadoop ecosystem who are closely connected with the Hadoop.
It is not a complete list. There are many more tools that even I don’t know. Maybe I
am drawing a complete map in the future.
48
16.3 Hadoop Ecosystem Components
Remember my big data platform blueprint? The blueprint has four stages: Ingest, store,
analyse and display.
Because of the Hadoop ecosystem” the different tools in these stages can work together
perfectly.
Here’s an example:
You use Apache Kafka to ingest data, and store the it in HDFS. You do the analytics
with Apache Spark and as a backend for the display you store data in Apache HBase.
To have a working system you also need YARN for resource management. You also need
Zookeeper, a configuration management service to use Kafka and HBase
As you can see in the picture below each project is closely connected to the other.
Spark for instance, can directly access Kafka to consume messages. It is able to access
HDFS for storing or processing stored data.
It also can write into HBase to push analytics results to the front end.
49
The cool thing of such ecosystem is that it is easy to build in new functions.
Want to store data from Kafka directly into HDFS without using Spark?
No problem, there is a project for that. Apache Flume has interfaces for Kafka and
HDFS.
It can act as an agent to consume messages from Kafka and store them into HDFS. You
even do not have to worry about Flume resource management.
Flume can use Hadoop’s YARN resource manager out of the box.
Although Hadoop is so popular it is not the silver bullet. It isn’t the tool that you should
use for everything.
Often times it does not make sense to deploy a Hadoop cluster, because it can be overkill.
Hadoop does not run on a single server.
You basically need at least five servers, better six to run a small cluster. Because of that.
the initial platform costs are quite high.
50
One option you have is to use a specialized systems like Cassandra, MongoDB or other
NoSQL DB’s for storage. Or you move to Amazon and use Amazon’s Simple Storage
Service, or S3.
Guess what the tech behind S3 is. Yes, HDFS. That’s why AWS also has the equivalent
to MapReduce named Elastic MapReduce.
The great thing about S3 is that you can start very small. When your system grows you
don’t have to worry about s3’s server scaling.
Yes, I definitely recommend you to get to now how Hadoop works and how to use it. As
I have shown you in this article, the ecosystem is quite large.
Many big data projects use Hadoop or can interface with it. Thats why it is generally a
good idea to know as many big data technologies as possible.
Not in depth, but to the point that you know how they work and how you can use them.
Your main goal should be to be able to hit the ground running when you join a big data
project.
Plus, most of the technologies are open source. You can try them out for free.
51
17 Docker
Have you played around with Docker yet? If you’re a data science learner or a data
scientist you need to check it out!
It’s awesome because it simplifies the way you can set up development environments for
data science. If you want to set up a dev environment you usually have to install a lot of
packages and tools.
What this does is you basically mess up your operating system. If you’re a starter you
don’t know which packages you need to install. You don’t know which tools you need to
install.
If you want to for instance start with Jupyter notebooks you need to install that on your
PC somehow. Or you need to start installing tools like PyCharmor Anaconda.
All that gets added to your system and so you mess up your system more and more
and more. What Docker brings you, especially if you’re on a Mac or a Linux system is
simplicity.
Because it is so easy to install on those systems. Another cool thing about docker images
is you can just search them in the Docker store, download them and install them on your
system.
52
You get a list of images you can download. You download one, start it up, you go to the
browser hit up the URL and just start coding.
Start doing the work. The only other thing you need to do is bind some drives to that
instance so you can exchange files. And then that’s it!
There is no way that you can crash or mess up your system. It’s all encapsulated into
Docker.Why this works is because Docker has natively access to your hardware.
It’s not a completely virtualized environment like a VirtualBox. An image has the upside
that you can take it wherever you want. So if your on your PC at home use that there.
Make a quick build, take the image and go somewhere else. Install the image which is
usually quite fast and just use it like you’re at home.
I am getting into Docker a lot more myself. For a bit different reasons.
What I’m looking for is using Docker with Kubernetes. With Kubernetes you can auto-
mate the whole container deployment process.
The idea with is that you have a cluster of machines. Lets say you have 10 server cluster
and you run Kubernetes on them.
Kubernetes lets you spin up Docker containers on-demand to execute tasks. You can set
up how much resources like CPU, RAM, Network, Docker container can use.
You can basically spin up containers, on the cluster on demand. When ever you need to
do a analytics task.
53
17.3 How to create, start,stop a Container
17.5 Kubernetes
Podcast about how data science learners use Docker (for data scientists): https://anchor.fm/andreaskayy
Data-Science-Go-Docker-e10n7u
Create a container:
d o c k e r run CONTAINER −−network NETWORK
Inspect the container configuration. For instance network settings and so on:
d o c k e r i n s p e c t CONTAINER
54
Create a new network:
d o c k e r network c r e a t e NETWORK −−d r i v e r b r i d g e
Remove a network
network rm NETWORK
55
18 REST APIs
APIs, Application Programming Interfaces are the cornerstones of any great data plat-
form.
In this podcast episode we look into the Twitter API. It’s a great example how to build
an API
Podcast Episode: #081 Twitter API Research Data Engineering Course Part 5
In this episode we look into the Twitter API documentation, which I love by the
way. How can we get old tweets for a certain hashtags and how to get current live
tweets for these hashtags?
Audio Click here to listen
Youtube Click here to watch
Jersey:
https://jersey.github.io/documentation/latest/getting-started.html
Swagger:
56
https://github.com/swagger-api/swagger-core/wiki/Swagger-2.X---Getting-started
Jersey vs Swagger:
https://stackoverflow.com/questions/36997865/what-is-the-difference-between-swagger-api-and-jax-rs
Spring Framework:
https://spring.io/
https://stackoverflow.com/questions/26824423/what-is-the-difference-among-spring-rest-service-and-je
57
19 Databases
19.1.1 PostgreSQL DB
Homepage:
https://www.postgresql.org/
Postgre vs Mongodb:
https://blog.panoply.io/postgresql-vs-mongodb
The Hadoop distributed file system, or HDFS, allows you to store files in Hadoop. The
difference between HDFS and other file systems like NTFS or EXT is that it is a dis-
tributed one.
58
Podcast Episode: #056 NoSQL Key Value Stores Explained with HBase
What is the difference between SQL and NoSQL? In this episode I show you on the
example of HBase how a key/value store works.
YouTube Click here to watch
Audio Click here to listen
Table 19.1: Podcast: 056 NoSQL Key Value Stores Explained with HBase
A typical file system stores your data on the actual hard drive. It is hardware dependent.
If you have two disks then you need to format every disk with its own file system. They
are completely separate.
You then decide on which disk you physically store your data.
Not only does it span over many disks in a server. It also spans over many servers.
HDFS will automatically place your files somewhere in the Hadoop server collective.
It will not only store your file, Hadoop will also replicate it two or three times (you can
define that). Replication means replicas of the file will be distributed to different servers.
This gives you superior fault tolerance. If one server goes down, then your data stays
available on a different server.
Another great thing about HDFS is, that there is no limit how big the files can be. You
can have server log files that are terabytes big.
How can files get so big? HDFS allows you to append data to files. Therefore, you can
continuously dump data into a single file without worries.
HDFS physically stores files different then a normal file system. It splits the file into
blocks.
These blocks are then distributed and replicated on the Hadoop cluster. The splitting
happens automatically.
In the configuration you can define how big the blocks should be. 128 megabyte or 1
gigabyte?
No problem at all.
This mechanic of splitting a large file in blocks and distributing them over the servers is
great for processing. See the MapReduce section for an example.
59
Figure 19.1: HDFS Master and Data Nodes
60
19.2.3 Document Store MongoDB
Links:
What is mongoDB:
https://www.guru99.com/what-is-mongodb.html#4
https://www.mongodb.com/what-is-mongodb
https://en.wikipedia.org/wiki/BSON
https://www.mkyong.com/mongodb/mongodb-hello-world-example
https://dzone.com/articles/real-time-analytics-on-mongodb-data-in-power-bi
https://www.mongodb.com/scale/when-to-use-apache-spark-with-mongodb
https://blog.timescale.com/how-to-store-time-series-data-mongodb-vs-timescaledb-postgresql-a739397
http://www.sarahmei.com/blog/2013/11/11/why-you-should-never-use-mongodb/
MongoDB vs Cassandra:
https://blog.panoply.io/cassandra-vs-mongodb
61
19.2.4 Elasticsearch Search Engine and Document Store
Elasticsearch is not a DB but firstly a search engine that indexes JSON documents.
Links:
JSONs need to be flattened heres how to work with nested objects in the JSON:
https://www.elastic.co/guide/en/elasticsearch/reference/current/nested.html
Indexing basics:
https://www.slideshare.net/knoldus/deep-dive-into-elasticsearch
62
Benchmarks how fast Elasticsearch is:
https://medium.appbase.io/benchmarking-elasticsearch-1-million-writes-per-sec-bf37e7ca8a4c
19.2.6 Impala
19.2.7 Kudu
Key concepts:
https://docs.influxdata.com/influxdb/v1.7/concepts/key concepts/
https://towardsdatascience.com/processing-time-series-data-in-real-time-with-influxdb-and-structured
https://medium.com/@xaviergeerinck/building-a-real-time-streaming-dashboard-with-spark-grafana-c
63
Performance Dashboard Spark and InfluxDB:
https://db-blog.web.cern.ch/blog/luca-canali/2019-02-performance-dashboard-apache-spark
Other alternatives for time series databases are: DalmatinerDB, InfluxDB, Prometheus,
Riak TS, OpenTSDB, KairosDB
64
20 Data Processing and Analytics -
Frameworks
Podcast Episode: #039 Is ETL Dead For Data Science & Big Data?
Is ETL dead in Data Science and Big Data? In today’s podcast I share with you
my views on your questions regarding ETL (extract, transform, load). Is ETL still
practiced or did pre processing & cleansing replace it What would replace ETL in
Data Engineering
YouTube Click here to watch
Audio Click here to listen
Table 20.1: Podcast: 039 Is ETL Dead for Data Science and Big Data?
Today’s topic are the different methods of streaming: At most once, at least once and
exactly once.
What this means and why it is so important to keep them in mind when creating a
solution. That is what you will find out in this article.
65
20.2.2 At Least Once
At least once, means a message gets processed in the system once or multiple times. So
with at least once it’s not possible that a message gets into the system and is not getting
processed.
One example where at least once processing can be used is when you think about a fleet
management of cars. You get GPS data from cars and that GPS data is transmitted with
a timestamp and the GPS coordinates.
It’s important that you get the GPS data at least once, so you know where the car is. If
you’re processing this data multiple times, it always has the the timestamp with it.
Because of that it does not matter that it gets processed multiple times, because of the
timestamp. Or that it would be stored multiple times, because it would just override the
existing one.
The second streaming method is at most once. At most once means that it’s okay to
drop some information, to drop some messages.
But it’s important that a message is only only processed once as a maximum.
A example for this is event processing. Some event is happening and that event is not
important enough, so it can be dropped. It doesn’t have any consequences when it gets
dropped.
But when that event happens it’s important that it does not get processed multiple times.
Then it would look as if the event happened five or six times instead of only one.
Think about engine misfires. If it happens once, no big deal. But if the system tells you
it happens a lot you will think you have a problem with your engine.
Another thing is exactly once, this means it’s not okay to drop data, it’s not okay to lose
data and it’s also not okay to process one message what data said multiple times
66
A example for this is for instance banking. When you think about credit card transactions
it’s not okay to drop a transaction.
When dropped your payment is not going through. It’s also not okay to have a transaction
processed multiple times, because then you are paying multiple times.
All of this sounds very simple and logical. What kind of processing is done has to be a
requirement for your use case.
It needs to be thought about in the design process, because not every tool is supporting
all three methods. Very often you need to code your application very differently based
on the streaming method.
So, the tool of data processing needs to be chosen based on if you need exactly once, at
least once or if you need at most once.
20.3 MapReduce
Since the early days of the Hadoop eco system, the MapReduce framework is one of the
main components of Hadoop alongside the Hadoop file system HDFS.
Google for instance used MapReduce to analyse stored html content of websites through
counting all the html tags and all the words and combinations of them (for instance
headlines). The output was then used to create the page ranking for Google Search.
That was when everybody started to optimise his website for the google search. Serious
search engine optimisation was borne. That was the year 2004.
How MapReduce is working is, that it processes data in two phases: The map phase and
the reduce phase.
In the map phase, the framework is reading data from HDFS. Each dataset is called an
input record.
Then there is the reduce phase. In the reduce phase, the actual computation is done
and the results are stored. The storage target can either be a database or back HDFS or
something else.
67
After all it’s Java – so you can implement what you like.
The magic of MapReduce is how the map and reduce phase are implemented and how
both phases are working together.
The map and reduce phases are parallelised. What that means is, that you have multiple
map phases (mappers) and reduce phases (reducers) that can run in parallel on your
cluster machines.
Here’s an example how such a map and reduce process works with data:
First of all, the whole map and reduce process relies heavily on using key/value pairs.
That’s what the mappers are for.
68
In the map phase input data, for instance a file, gets loaded and transformed into key/-
value pairs.
When each map phase is done it sends the created key/value pairs to the reducers where
they are getting sorted by key. This means, that an input record for the reduce phase is
a list of values from the mappers that all have the same key.
Then the reduce phase is doing the computation of that key and its values and outputting
the results.
How many mappers and reducers can you use in parallel? The number of parallel map
and reduce processes depends on how many CPU cores you have in your cluster. Every
mapper and every reducer is using one core.
This means that the more CPU cores you actually have, the more mappers you can use,
the faster the extraction process can be done. The more reducers you are using the faster
the actual computation is being done.
20.3.2 Example
As I said before, MapReduce works in two stages, map and reduce. Often these stages
are explained with a word count task.
Personally, I hate this example because counting stuff is to trivial and does not really
show you what you can do with MapReduce. Therefore, we are going to use a more real
world use-case from the world of the internet of things (IoT).
IoT applications create an enormous amount of data that has to be processed. This data
is generated by physical sensors who take measurements, like room temperature at 8.00
o’Clock.
Every measurement consists of a key (the timestamp when the measurement has been
taken) and a value (the actual value measured by the sensor).
Because you usually have more than one sensor on your machine, or connected to your
system, the key has to be a compound key. Compound keys contain additionally to the
measurement time information about the source of the signal.
But, let’s forget about compound keys for now. Today we have only one sensor. Each
measurement outputs key/value pairs like: Timestamp-Value.
The goal of this exercise is to create average daily values of that sensor’s data.
69
The image below shows how the map and reduce process works.
First, the map stage loads unsorted data (input records) from the source (e.g. HDFS) by
key and value (key:2016-05-01 01:02:03, value:1).
Then, because the goal is to get daily averages, the hour:minute:second information is
cut from the timestamp.
After all parallel map phases are done, each key/value pair gets sent to the one reducer
who is handling all the values for this particular key.
Every reducer input record then has a list of values and you can calculate (1+5+9)/3,
(2+6+7)/3 and (3+4+8)/3. That’s all.
Yes, you need to cut the key differently. You then would need to cut it like this: “2016-
05-01 01:02”. Keeping the Hour and minute information in the key.
What you can also see is, why map reduce is so great for doing parallel work. In this case,
the map stage could be done by nine mappers in parallel because each map is independent
from all the others.
The reduce stage could still be done by three tasks in parallel. One each for orange, blue
and one for green.
That means, if your dataset would be 10 times as big and you’d have 10 times the
machines, the time to do the calculation would be the same.
70
20.3.3 What is the limitation of MapReduce? – available
MapReduce is awesome for simpler analytics tasks, like counting stuff. It just has one
flaw: It has only two stages Map and Reduce.
First MapReduce loads the data from HDFS into the mapping function. There you
prepare the input data for the processing in the reducer. After the reduce is finished the
results get written to the data store.
The problem with MapReduce is that there is no simple way to chain multiple map and
reduce processes together. At the end of each reduce process the data must be stored
somewhere.
This fact makes it very hard to do complicated analytics processes. You would need to
chain MapReduce jobs together.
Chaining jobs with storing and loading intermediate results just makes no sense.
Another issue with MapReduce is that it is not capable of streaming analytics. Jobs take
some time to spin up, do the analytics and shut down. Basically Minutes of wait time
are totally normal.
This is a big negative point in a more and more real time data processing world.
I talked about the three methods of data streaming in this podcast: https://anchor.fm/andreaskayy/emb
Methods-of-Streaming-Data-e15r6o
Spark is a complete in memory framework. Data gets loaded from, for instance hdfs, into
the memory of workers.
There is no longer a fixed map and reduce stage. Your code can be as complex as you
want.
71
Once in memory, the input data and the intermediate results stay in memory (until the
job finishes). They do not get written to a drive like with MapReduce.
This makes Spark the optimal choice for doing complex analytics. It allows you for
instance to do Iterative processes. Modifying a dataset multiple times in order to create
an output is totally easy.
Streaming analytics capability is also what makes spark so great. Spark has natively the
option to schedule a job to run every X seconds or X milliseconds.
As a result, Spark can deliver you results from streaming data in “real time”.
There are some very misleading articles out there titled Spark or Hadoop, Spark is better
than Hadoop or even Spark is replacing Hadoop.
So, it’s time to show you the differences between Spark and Hadoop. After this you will
know when and for what you should use Spark and Hadoop.
You’ll also understand why Hadoop or Spark is the totally wrong question.
To make it clear how Hadoop differs from Spark I created this simple feature table:
Hadoop is used to store data in the Hadoop Distributed File System (HDFS). It can
analyse the stored data with MapReduce and manage resources with YARN.
However, Hadoop is more than just storage, analytics and resource management. There’s
a whole eco system of tools around the Hadoop core. I’ve written about tis eco system
in this article: What is Hadoop and why is it so freakishly popular. You should check it
out as well.
72
Compared to Hadoop, Spark is “just” an analytics framework. It has no storage capa-
bility. Although it has a standalone resource management, you usually don’t use that
feature.
So, if Hadoop and Spark are not the same things, can they work together?
Absolutely! Here’s how the first picture will look if you combine Hadoop with Spark:
As Storage you use the Hadoop distributed file system. Analytics is done with Apache
Spark and Yarn is taking care of the resource management.
From a platform architecture perspective, Hadoop and Spark are usually managed on the
same cluster. This means on each server where a HDFS data node is running, a spark
worker thread runs as well.
Spark is able to determine on which data node the needed data is stored. This allows a
direct load of the data from the local storage into the memory of the machine.
You need to make sure that your physical resources are distributed perfectly between
the services. This is especially the case when you run Spark workers with other Hadoop
services on the same machine.
It just would not make sense to have two resource managers managing the same server’s
resources. Sooner or later they will get in each others way.
So, the question is not Spark or Hadoop. The question has to be: Should you use Spark
or MapReduce alongside Hadoop’s HDFS and YARN.
73
20.4.6 My simple rule of thumb:
If you are doing simple batch jobs like counting values or doing calculating averages: Go
with MapReduce.
If you need more complex analytics like machine learning or fast stream processing: Go
with Apache Spark.
Spark jobs can be programmed in a variety of languages. That makes creating analytic
processes very user-friendly for data scientists.
Spark supports Python, Scala and Java. With the help of SparkR you can even connect
your R program to a Spark cluster.
If you are a data scientist who is very familiar with Python just use Python, its great. If
you know how to code Java I suggest you start using Scala.
Spark jobs are easier to code in Scala than in Java. In Scala you can use anonymous
functions to do processing.
With Java 8 simplified function calls were introduced with lambda expressions. Still, a
lot of people, including me prefer Scala over Java.
Another thing is data locality. I always make the point, that processing data locally
where it is stored is the most efficient thing to do.
That’s exactly what Spark is doing. You can and should run Spark workers directly on
the data nodes of your Hadoop cluster.
Spark can then natively identify on what data node the needed data is stored. This
enables Spark to use the worker running on the machine where the data is stored to load
74
the data into the memory.
The downside of this setup is that you need more expensive servers. Because Spark
processing needs stronger servers with more RAM and CPUs than a “pure” Hadoop
setup.
20.4.16 MLlib:
The machine learning library MLlib is included in Spark so there is often no need to
import another library.
I have to admit because I am not a data scientist I am not an expert in machine learning.
75
From what I have seen and read though the machine learning framework MLlib is a nice
treat for data scientists wanting to train and apply models with Spark.
From a solution architect’s point of view Spark is a perfect fit for Hadoop big data
platforms. This has a lot to do with cluster deployment and management.
Companies like Cloudera, MapR or Hortonworks include Spark into their Hadoop distri-
butions. Because of that, Spark can be deployed and managed with the clusters Hadoop
management web fronted.
This makes the process for deploying and configuring a Spark cluster very quick and
admin friendly.
When running a computing framework you need resources to do computation: CPU time,
RAM, I/O and so on. Out of the box Spark can manage resources with it’s stand-alone
resource manager.
If Spark is running in an Hadoop environment you don’t have to use Spark’s own stand-
alone resource manager. You can configure Spark to use Hadoop’s YARN resource man-
agement.
Why would you do that? It allows YARN to efficiently allocate resources to your Hadoop
and Spark processes.
Having a single resource manager instead of two independent ones makes it a lot easier
to configure the resource management.
20.6 StreamSets
https://youtu.be/djt8532UWow
76
Figure 20.6: Spark Resource Management With YARN
https://www.youtube.com/watch?v=Qm5e574WoCU&t=2s
https://github.com/gschmutz/stream-processing-workshop/tree/master/04-twitter-data-ingestion-with
https://streamsets.com/blog/streaming-data-twitter-analysis-spark/
77
21 Apache Kafka
78
−e ALLOW PLAINTEXT LISTENER=y e s \
bitnami / k a f k a : l a t e s t
79
22 Machine Learning
Machine learning in production is using stream and batch processing. In the batch
processing layer you are creating the models, because you have all the data available for
training.
In the stream in processing layer you are using the created models, you are applying them
to new data.
The idea that you need to incorporate is that it is a constant cycle. Training, applying,
re-training, pushing into production and applying.
What you don’t want to do is, you don’t want to do this manually. You need to figure out
a process of automatic retraining and automatic pushing to into production of models.
In the retraining phase the system automatically evaluates the training. If the model no
longer fits it works as long as it needs to create a good model.
After the evaluation of the model is complete and it’s good, the model gets pushed into
production. Into the stream processing.
80
22.4 Why machine learning in production is harder
then you think – available
How to automate machine learning is something that drives me day in and day out.
What you do in development or education is, that you create a model and fit it to the
data. Then that model is basically done forever.
Where I’m coming from, the IoT world, the problem is that machines are very different.
They behave very different and experience wear.
Machines have certain processes that decrease the actual health of the machine. Machine
wear is a huge issue. Models that that are built on top of a good machine don’t work
forever.
When the Machine wears out, the models need to be adjusted. They need to be main-
tained, retrained.
Automatic re-training and re-deploying is a very big issue, a very big problem for a lot of
companies. Because most existing platforms don’t have this capability (I actually haven’t
seen one until now).
Look at AWS machine learning for instance. The process is: build, train, tune deploy.
Where’s the loop of retraining?
You can create models and then use them in production. But this loop is almost nowhere
to be seen.
It is a very big issue that needs to be solved. If you want to do machine learning in
production you can start with manual interaction of the training, but at some point you
need to automate everything.
81
22.7 Training Parameter Management
Take deep learning for instance. To train you are manipulating for instance:
How many layers do you use. The depth of the layers, which means how many neurons
you have in a layer. What activation function you use, how long are you training and so
on.
You also need to keep track of what data you used to train which model.
All those parameters need to be manipulated automatically, models trained and tested.
To do all that, you basically need a database that keeps track of those variables.
How to automate this, for me, is like the big secret. I am still working on figuring it out.
Did you already have the problem of automatic re-training and deploying of models as
well?
Were you able to use a cloud platform like Google, AWS or Azure?
Many people still are not convinced that machine learning works reliably. But they want
analytics insight and most of the time machine learning is the way to go.
This means, when you are working with customers you need to do a lot of convincing.
Especially if they are not into machine learning themselves.
82
22.10 No Rules, No Physical Models
Many people are still under the impression that analytics only works when it’s based on
physics. When there are strict mathematical rules to a problem.
“Sere has to be a Rule for Everysing!” (imagine a German accent) When you’re engi-
neering you are calculating stuff based on physics and not based on data. If you are
constructing an airplane wing, you better make sure to use calculations so it doesn’t fall
off.
Machine learning has been around for decades. It didn’t quite work as good as people
hoped. We have to admit that. But there is this preconception that it still doesn’t work.
Somehow you need to convince people that it is a viable approach. That learning from
data to make predictions is working perfectly.
As a data scientist you have one ace up your sleeve, it’s the obvious one:
You can use that data and those statistics to counter peoples preconceptions. It’s very
powerful if someone says: “This doesn’t work”
You bring the data. You show the statistics and you show that it works reliably.
Data doesn’t lie. You can’t fight data. The data is always right.
83
22.12 Data is Stronger Than Opinions
This is also why I believe that autonomous driving will come quicker than many of us
think. Because a lot of people say, they are not safe. That you cannot rely on those cars.
The thing is: When you have the data you can do the statistics.
You can show people that autonomous driving really works reliably. You will see, the
question of: Is this is this allowed or is this not allowed? Will be gone quicker than you
think.
Because government agencies can start testing the algorithms based on predefined sce-
narios. They can run benchmarks and score the cars performance.
The motor agency has the statistics. The stats show people how good cars work.
Companies like Tesla, they have it very easy. Because the data is already there.
They just need to show us that the algorithms work. The end.
Link to the OLX Slideshare with pros, cons and how to use Sagemaker: https://www.
slideshare.net/mobile/AlexeyGrigorev/image-models-infrastructure-at-olx
84
23 Data Visualization
This section does not contain any text that’s why the page is messed up
85
23.3.1 Tomcat
23.3.2 Jetty
23.3.3 NodeRED
23.3.4 React
23.4.1 Tableau
23.4.2 PowerBI
23.4.3 Quliksense
86
Part III
87
24 What We Want To Do
• Twitter data to predict best time to post using the hashtag datascience or ai
• Top users
88
25 Thoughts On Choosing A
Development Environment
For a local environment you need a good PC. I thought a bit about a budget build around
1.000 Dollars or Euros.
89
26 A Look Into the Twitter API
90
27 Ingesting Tweets with Apache
Nifi
Podcast Episode: #082 Reading Tweets With Apache Nifi & IaaS vs PaaS vs SaaS
In this podcast we are trying to read Twitter Data with Nifi
YouTube Click here to watch
Table 27.2: Podcast: 085 Trying to read Tweets with Nifi Part 2
91
28 Writing from Nifi to Apache
Kafka
Table 28.1: Podcast: 086 How to Write from Nifi to Kafka Part 1
Table 28.2: Podcast: 088 How to Write from Nifi to Kafka Part 2
92
29 Apache Zeppelin
d o c k e r run −d −p 8 0 8 1 : 8 0 8 0 −−rm \
−v / U s e r s / xxxx /Documents/ D o c k e r F i l e s / l o g s : / l o g s \
−v / U s e r s / xxxx /Documents/ D o c k e r F i l e s / Notebooks : / notebook \
−e ZEPPELIN LOG DIR=’/ l o g s ’ \
−e ZEPPELIN NOTEBOOK DIR=’/ notebook ’ \
−−network app−t i e r −−name z e p p e l i n apache / z e p p e l i n : 0 . 7 . 3
93
30 Switch Processing from Zeppelin
to Spark
94
Part IV
Case Studies
95
31 How I do Case Studies
Slides:
https://medium.com/airbnb-engineering/airbnb-engineering-infrastructure/home
https://www.datasciencecentral.com/profiles/blogs/20-data-science-systems-used-by-amazon-to-opera
https://aws.amazon.com/solutions/case-studies/amazon-migration-analytics/
https://www.slideshare.net/databricks/spark-sql-adaptive-execution-unleashes-the-power-of-cluster-in
96
31.3 Data Science @Blackrock
https://www.slideshare.net/DataStax/maintaining-consistency-across-data-centers-randy-fradin-black
Big Data in der Automobilindustrie – Daten aus dem Fahrzeug nutzen https://www.
unibw.de/code.../ws3 bigdata vortrag widmann.pdf
Slides:
https://www.slideshare.net/ConfluentInc/data-streaming-ecosystem-management-at-bookingcom?
ref=https://www.confluent.io/kafka-summit-sf18/data-streaming-ecosystem-management
https://www.slideshare.net/SparkSummit/productionizing-behavioural-features-for-machine-learning-w
https://www.slideshare.net/ConfluentInc/data-streaming-ecosystem-management-at-bookingcom?
ref=https://www.confluent.io/kafka-summit-sf18/data-streaming-ecosystem-management
Druid: https://towardsdatascience.com/introduction-to-druid-4bf285b92b5a
Slides:
http://www.lhc-facts.ch/index.php?page=datenverarbeitung
97
Podcast Episode: #065 Data Engineering At CERN Case Study
How is CERN doing Data Engineering? The must get huge amounts of data from
the Large Hydron Colider. Let’s check it out.
YouTube Click here to watch
Audio Click here to listen
https://openlab.cern/sites/openlab.web.cern.ch/files/2018-05/kubeconeurope2018-cern-180507122303.
pdf
https://www.slideshare.net/SparkSummit/next-cern-accelerator-logging-service-with-jakub-wozniak
https://databricks.com/session/the-architecture-of-the-next-cern-accelerator-logging-service
http://opendata.cern.ch
https://gobblin.apache.org
https://www.slideshare.net/databricks/cerns-next-generation-data-analysis-platform-with-apache-spar
https://www.slideshare.net/SparkSummit/realtime-detection-of-anomalies-in-the-database-infrastruct
https://medium.com/disney-streaming/delivering-data-in-real-time-via-auto-scaling-kinesis-streams-7
https://berlin-2017.flink-forward.org/kb sessions/drivetribes-kappa-architecture-with-apache-flink/
https://www.slideshare.net/FlinkForward/flink-forward-berlin-2017-aris-kyriakos-koliopoulos-drivetrib
https://blogs.dropbox.com/tech/2019/01/finding-kafkas-throughput-limit-in-dropbox-infrastructure/
98
31.10 Data Science @Ebay
https://www.slideshare.net/databricks/moving-ebays-data-warehouse-over-to-apache-spark-spark-as-c
https://www.slideshare.net/databricks/analytical-dbms-to-apache-spark-auto-migration-framework-wi
https://www.slideshare.net/BrandonOBrien/spark-streaming-kafka-best-practices-w-brandon-obrien
https://www.slideshare.net/Naveen1914/brandon-obrien-streamingdata
https://code.fb.com/core-data/apache-spark-scale-a-60-tb-production-use-case/
http://www.unofficialgoogledatascience.com/
https://ai.google/research/teams/ai-fundamentals-applications/
https://cloud.google.com/solutions/big-data/
https://datafloq.com/read/google-applies-big-data-infographic/385
https://www.slideshare.net/databricks/building-a-versatile-analytics-pipeline-on-top-of-apache-spark-
https://sf-2017.flink-forward.org/kb sessions/streaming-models-how-ing-adds-models-at-runtime-to-ca
https://www.slideshare.net/SparkSummit/lessons-learned-developing-and-managing-massive-300tb-ap
99
31.17 Data Science @LinkedIn
Slides:
https://engineering.linkedin.com/teams/data#0
https://www.slideshare.net/yaelgarten/building-a-healthy-data-ecosystem-around-kafka-and-hadoop-l
https://engineering.linkedin.com/teams/data/projects/pinot
https://pinot.readthedocs.io/en/latest/intro.html#
https://towardsdatascience.com/building-machine-learning-at-linkedin-scale-f08bd9a63f0a
http://samza.apache.org
https://www.slideshare.net/ConfluentInc/more-data-more-problems-scaling-kafkamirroring-pipelines-a
ref=https://www.confluent.io/kafka-summit-sf18/more data more problems
https://www.slideshare.net/KhaiTran17/conquering-the-lambda-architecture-in-linkedin-metrics-platf
https://www.slideshare.net/Hadoop Summit/unified-batch-stream-processing-with-apache-samza
http://druid.io/docs/latest/design/index.html
https://eng.lyft.com/running-apache-airflow-at-lyft-6e53bb8fccff
Slides:
http://sites.nationalacademies.org/cs/groups/ssbsite/documents/webpage/ssb 182893.pdf
100
Podcast Episode: #067 Data Engineering At NASA Case Study
A look into how NASA is doing data engineering.
YouTube Click here to watch
Audio Click here to listen
http://www.socallinuxexpo.org/sites/default/files/presentations/OnSightCloudArchitecture-scale14x.
pdf
https://www.slideshare.net/SparkSummit/spark-at-nasajplchris-mattmann?qid=90968554-288e-454a-b
v=&b=&from search=4
Slides:
Netflix revolutionized how we watch movies and tv. Currently over 75 million users watch
125 million hours of Netflix content every day!
Netflix’s revenue comes from a monthly subscription service. So, the goal for Netflix is
to keep you subscribed and to get new subscribers.
To achieve this, Netflix is licensing movies from studios as well as creating its own original
movies and tv series.
But offering new content is not everything. What is also very important is, to keep you
watching content that already exists.
To be able to recommend you content, Netflix is collecting data from users. And it is
collecting a lot.
Currently, Netflix analyses about 500 billion user events per day. That results in a
stunning 1.3 petabytes every day.
101
All this data allows Netflix to build recommender systems for you. The recommenders
are showing you content that you might like, based on your viewing habits, or what is
currently trending.
The Netflix batch processing pipeline When Netflix started out, they had a very
simple batch processing system architecture.
The key components were Chuckwa, a scalable data collection system, Amazon S3 and
Elastic MapReduce.
Chuckwa wrote incoming messages into Hadoop sequence files, stored in Amazon S3.
These files then could be analysed by Elastic MapReduce jobs.
Netflix batch processing pipeline Jobs were executed regularly on a daily and hourly
basis. As a result, Netflix could learn how people used the services every hour or once a
day.
Know what customers want: Because you are looking at the big picture you can
create new products. Netflix uses insight from big data to create new tv shows and
movies.
They created House of Cards based on data. There is a very interesting Ted talk about
this you should watch:
Batch processing also helps Netflix to know the exact episode of a TV show that gets you
hooked. Not only globally but for every country where Netflix is available.
102
They know exactly what show works in what country and what show does not.
It helps them create shows that work in everywhere or select the shows to license in
different countries. Germany for instance does not have the full library that Americans
have :(
We have to put up with only a small portion of tv shows and movies. If you have to
select, why not select those that work best.
Batch processing is not enough As a data platform for generating insight the
Cuckwa pipeline was a good start. It is very important to be able to create hourly
and daily aggregated views for user behavior.
The only problem is: With batch processing you are basically looking into the past.
For Netflix, and data driven companies in general, looking into the past is not enough.
They want a live view of what is happening.
The trending now feature One of the newer Netflix features is “Trending now”. To
the average user it looks like that “Trending Now” means currently most watched.
This is what I get displayed as trending while I am writing this on a Saturday morning
at 8:00 in Germany. But it is so much more.
What is currently being watched is only a part of the data that is used to generate
“Trending Now”.
“Trending now” is created based on two types of data sources: Play events and Impression
events.
What messages those two types actually include is not really communicated by Netflix.
I did some research on the Netflix Techblog and this is what I found out:
Play events include what title you have watched last, where you did stop watching, where
you used the 30s rewind and others. Impression events are collected as you browse the
103
Netflix Library like scroll up and down, scroll left or right, click on a movie and so on
Basically, play events log what you do while you are watching. Impression events are
capturing what you do on Netflix, while you are not watching something.
Netflix real-time streaming architecture Netflix uses three internet facing services
to exchange data with the client’s browser or mobile app. These services are simple
Apache Tomcat based web services.
The service for receiving play events is called “Viewing History”. Impression events are
collected with the “Beacon” service.
The “Recommender Service” makes recommendations based on trend data available for
clients.
Messages from the Beacon and Viewing History services are put into Apache Kafka. It
acts as a buffer between the data services and the analytics.
Beacon and Viewing History publish messages to Kafka topics. The analytics subscribes
to the topics and gets the messages automatically delivered in a first in first out fashion.
After the analytics the workflow is straight forward. The trending data is stored in a
Cassandra Key-Value store. The recommender service has access to Cassandra and is
making the data available to the Netflix client.
The algorithms how the analytics system is processing all this data is not known to the
public. It is a trade secret of Netflix.
What is known, is the analytics tool they use. Back in Feb. 2015 they wrote in the tech
blog that they use a custom made tool.
104
They also stated, that Netflix is going to replace the custom made analytics tool with
Apache Spark streaming in the future. My guess is, that they did the switch to Spark
some time ago, because their post is more than a year old.
Slides:
https://www.slideshare.net/mobile/AlexeyGrigorev/image-models-infrastructure-at-olx
https://www.slideshare.net/SparkSummit/spark-summit-eu-talk-by-sebastian-schroeder-and-ralf-sigm
https://www.paypal-engineering.com/tag/data/
Slides:
105
https://www.slideshare.net/ConfluentInc/pinterests-story-of-streaming-hundreds-of-terabytes-of-pins-
ref=https://www.confluent.io/kafka-summit-sf18/pinterests-story-of-streaming-hundreds-of-terabytes
https://www.slideshare.net/ConfluentInc/building-pinterest-realtime-ads-platform-using-kafka-stream
ref=https://www.confluent.io/kafka-summit-sf18/building-pinterest-real-time-ads-platform-using-kafka
https://medium.com/@Pinterest Engineering/building-a-real-time-user-action-counting-system-for-ad
https://medium.com/pinterest-engineering/goku-building-a-scalable-and-high-performant-time-series-
https://medium.com/pinterest-engineering/building-a-dynamic-and-responsive-pinterest-7d410e99f0a9
https://medium.com/@Pinterest Engineering/building-pin-stats-25ec8460e924
https://medium.com/@Pinterest Engineering/improving-hbase-backup-efficiency-at-pinterest-86159da
https://medium.com/@Pinterest Engineering/pinterest-joins-the-cloud-native-computing-foundation-e
https://medium.com/@Pinterest Engineering/using-kafka-streams-api-for-predictive-budgeting-9f58d2
https://medium.com/@Pinterest Engineering/auto-scaling-pinterest-df1d2beb4d64
https://engineering.salesforce.com/building-a-scalable-event-pipeline-with-heroku-and-salesforce-2549c
Table 31.9: Podcast: 059 What Is The Siemens Mindsphere IoT Platform?
https://speakerdeck.com/vananth22/streaming-data-pipelines-at-slack
106
31.28 Data Science @Spotify
Slides:
https://labs.spotify.com/2016/02/25/spotifys-event-delivery-the-road-to-the-cloud-part-i/
https://labs.spotify.com/2016/03/03/spotifys-event-delivery-the-road-to-the-cloud-part-ii/
https://labs.spotify.com/2016/03/10/spotifys-event-delivery-the-road-to-the-cloud-part-iii/
https://www.slideshare.net/InfoQ/scaling-the-data-infrastructure-spotify
https://www.datanami.com/2018/05/16/big-data-file-formats-demystified/
https://labs.spotify.com/2017/04/26/reliable-export-of-cloud-pubsub-streams-to-cloud-storage/
https://labs.spotify.com/2017/11/20/autoscaling-pub-sub-consumers/
https://www.slideshare.net/planetcassandra/symantec-cassandra-data-modelling-techniques-in-action
https://www.slideshare.net/databricks/scalable-monitoring-using-apache-spark-and-friends-with-utkar
Slides:
https://www.slideshare.net/sawjd/real-time-processing-using-twitter-heron-by-karthik-ramasamy
https://www.slideshare.net/sawjd/big-data-day-la-2016-big-data-track-twitter-heron-scale-karthik-ram
107
Podcast Episode: #072 Data Engineering At Twitter Case Study
How is Twitter doing Data Engineering? Oh man, they have a lot of cool things to
share these tweets.
YouTube Click here to watch
Audio Click here to listen
https://techjury.net/stats-about/twitter/
https://developer.twitter.com/en/docs/tweets/post-and-engage/overview
https://www.slideshare.net/prasadwagle/extracting-insights-from-data-at-twitter
https://blog.twitter.com/engineering/en us/topics/insights/2018/twitters-kafka-adoption-story.
html
https://blog.twitter.com/engineering/en us/topics/infrastructure/2017/the-infrastructure-behind-twitt
html
https://blog.twitter.com/engineering/en us/topics/infrastructure/2019/the-start-of-a-journey-into-the
html
https://www.slideshare.net/billonahill/twitter-heron-in-practice
https://streaml.io/blog/intro-to-heron
https://www.youtube.com/watch?v=3QHGhnHx5HQ
https://hbase.apache.org
https://db-engines.com/en/system/Amazon+DynamoDB%3BCassandra%3BGoogle+Cloud+
Bigtable%3BHBase
https://eng.uber.com/uber-big-data-platform/
https://eng.uber.com/aresdb/
https://www.uber.com/us/en/uberai/
108
31.33 Data Science @Upwork
https://www.slideshare.net/databricks/how-to-rebuild-an-endtoend-ml-pipeline-with-databricks-and-u
https://aws.amazon.com/de/blogs/big-data/our-data-lake-story-how-woot-com-built-a-serverless-data
Do me a favor and give these guys a follow on LinkedIn: LinkedIn of Michal: https:
//www.linkedin.com/in/michalgancarski/
Zalando has a tech blog with more infos and there is also a meetup in Berlin: Zalando
Blog: https://jobs.zalando.com/tech/blog/
109
Talk at Strata London slides: https://databricks.com/session/continuous-applications-at-scale-of-100-te
https://jobs.zalando.com/tech/blog/what-is-hardcore-data-science--in-practice/?gh src=
4n3gxh1
https://jobs.zalando.com/tech/blog/complex-event-generation-for-business-process-monitoring-using-a
110
Part V
111
Looking for a job or just want to know what people find important? In this chapter you
can find a lot of interview questions we collect on the stream.
Ultimately this should reach at least one thousand and one questions.
But Andreas, where are the answers?? Answers are for losers. I have been thinking
a lot about this and the best way for you to prepare and learn is to look into these
questions yourself.
This cookbook or Google will help you a long way. Some Questions we discuss directly
on the live stream.
112
32 Live Streams
113
33 All Interview Questions
The interview questions are roughly structured like the sections in the ”Basic data Engi-
neering Skills” part. This makes it easier to navigate this document. I still need to sort
them accordingly.
SQL DBs
• What is the difference between Clustered Index and Non-Clustered Index - with
examples?
The Cloud
• What is serverless
• How do you move from the ingest layer to the Cosumption layer? (In Serverless)
114
Linux
• What is crontab
Big Data
Kafka
• What is a topic
• How do you know if all messages in a topic have been fully consumed
• What is a producer
Coding
• Explain immutability
• What are AWS Lambda functions and why would you use them
115
NoSQL DBs
• What’s a columnstore
Hadoop
• What is HDFS
Lambda Architecture
• Can you sync the batch and streaming layer and if yes how
Python
116
• What is a data warehouse
APIs (REST)
• What is idempotency
Apache Spark
• What’s an RDD
• What is a dataframe
• What is a dataset
• What is Parquet
• What’s Avro
117
MapReduce
• What is a combiner
• What is a container
Data Pipelines
Airflow
• How to branch?
DataViszualization
• What’s a BI tool
118
Security/Privacy
• What is Kerberos
• What is a firewall
• Whats GDPR?
• What’s anonymization
Distrubuted Systems
• how clusters reach consensus (the answer was using consensus protocols like Paxos
or Raft). Good I didnt have to explain paxos
• What is the cap theorem / explain it (What factors should be considered when
choosing a DB?)
• How to choose right storage for different data consumers? It’s always a tricky
question
Apache Flink
• Flink vs Spark?
GitHub
Dev/Ops
119
• What is continuous deployment
• Difference CI/CD
Development / Agile
• What is Scrum
• What is OKR
120
Bibliography
[1] J. Ely and I. Stavrov1, Analyzing chalk dust and writing speeds: computational and
geometric approaches, BoDine Journal of Mathematics 3 (2001), 14-159.
121
List of Figures
122
List of Tables
2.1 Podcast: 050 Data Engineer Scientist or Analyst Which One Is For You? 12
2.2 Podcast: 048 From Wannabe Data Scientist To Engineer My Journey . . 14
10.1 Podcast: 082 Reading Tweets With Apache Nifi & IaaS vs PaaS vs SaaS 30
10.2 Podcast: 076 Cloud vs On-Premise . . . . . . . . . . . . . . . . . . . . . 30
16.1 Podcast: 060 What Is Hadoop And Is Hadoop Still Relevant In 2019? . . 47
19.1 Podcast: 056 NoSQL Key Value Stores Explained with HBase . . . . . . 59
19.2 Podcast: 093 What is MongoDB . . . . . . . . . . . . . . . . . . . . . . . 61
19.3 Podcast: What is Elasticsearch & Why is It So Popular? . . . . . . . . . 62
19.4 Podcast: Druid NoSQL DB And Analytics DB Introduction . . . . . . . 63
20.1 Podcast: 039 Is ETL Dead for Data Science and Big Data? . . . . . . . . 65
123
31.1 Podcast: 063 Data Engineering At Airbnb Case Study . . . . . . . . . . 96
31.2 Podcast: 064 Data Engineering At Booking.com Case Study . . . . . . . 97
31.3 Podcast: 065 Data Engineering At CERN Case Study . . . . . . . . . . . 98
31.4 Podcast: 073 Data Engineering At LinkedIn Case Study . . . . . . . . . 100
31.5 Podcast: 067 Data Engineering At NASA Case Study . . . . . . . . . . . 101
31.6 Podcast: 062 Data Engineering At Netflix Case Study . . . . . . . . . . . 101
31.7 Podcast: 083 Data Engineering at OLX Case Study . . . . . . . . . . . . 105
31.8 Podcast: 069 Engineering Culture At Pinterest . . . . . . . . . . . . . . . 105
31.9 Podcast: 059 What Is The Siemens Mindsphere IoT Platform? . . . . . . 106
31.10Podcast: 071 Data Engineering At Spotify Case Study . . . . . . . . . . 107
31.11Podcast: 072 Data Engineering At Twitter Case Study . . . . . . . . . . 108
31.12Podcast: 087 Data Engineering At Zalando Case Study . . . . . . . . . . 109
124