Big Data and Machine Learning

Fundamentals with Google Cloud Platform

 An explosion of data

“By 2020, some 50 billion smart

devices will be connected, along with
additional billions of smart sensors,
ensuring that the global supply of data
will continue to more than double
every two years”

https://www.mckinsey.com/business-functions/digital-mckinsey/our-insights/straight-talk-about-big-data
 An explosion of data

… and only about 1% of the data generated

today is actually analyzed

https://www.mckinsey.com/business-functions/digital-mckinsey/our-insights/straight-talk-about-big-data
 There is a great demand for data skills

Data Analyst Applied ML Engineer Tech Lead

Analyst Data Scientist Ethicist Analytics Manager

Statistician Social Scientist Decision Maker

Data Engineer
Applied ML
Data Engineer Researcher
Engineer
 Big Data Challenges

Migrating existing
Analyzing large
data workloads
datasets at scale
(ex: Hadoop, Spark jobs)

Building streaming Applying machine

data pipelines learning to your data
Agenda
Google Cloud Platform infrastructure
● Compute
● Storage
● Networking
● Security
Big data and ML products
● Google innovation timeline
● Choosing the right approach
Lab: Exploring Public Datasets in BigQuery
Activity: Explore a customer use case
 Seven cloud products with
Google’s mission one billion users

Organize the world’s information

and make it universally accessible
and useful.
 Big Data and ML Products

Compute Power Storage Networking

Security
 Big Data and ML Products

Compute Power Storage Networking

Security
Machine Learning Models require
significant compute resources
Shown: Automatic
Video Stabilization
for Google Photos

Data sources:
1. Image frames
(stills from video)
2. Phone gyroscope
3. Lens motion
 A single high-res image represents millions of data points to learn

← 3264px Width →

8 Megapixel resolution
←2448px Height →

3264 (w) x 2448 (h) x 3 (RGB) =

23,970,816
data points per image*

* More data = longer model training

times + more storage needed

3 “Layers” in depth for Red Blue Green

Google trains on its infrastructure and
deploys ML to phone hardware
 Build on Google infrastructure

This is what makes Google Google: its

physical network … and those many
thousands of servers that, in aggregate,
add up to the mother of all clouds.”
- Wired
Simply scaling the raw number of servers in Google’s
data centers isn’t enough

“If everyone spoke to their

phone for 3 minutes, we’d
exhaust all available
computing resources”

— Jeff Dean, 2014

 Will Moore’s Law save us?

https://cacm.acm.org/magazines/2018/9/230571-a-domain-specific-architecture-for-deep-neural-networks/fulltext
Tensor Processing Units (TPUs) are specialized ML hardware

Cloud TPU v2 Cloud TPU v3

180 teraflops 420 teraflops
64-GB High Bandwidth 128-GB HBM
Memory (HBM
TPUs enable faster models and more iterations

“Cloud TPU Pods have transformed

our approach to visual shopping by
delivering a 10X speedup over our
previous infrastructure.
+
We used to spend months training a
single image recognition model,
whereas now we can train much
more accurate models in a few
days on Cloud TPU Pods.

— Larry Colagiovanni
VP of New Product Development
 Creating a customizable virtual machine on Google Cloud

Create with the Google Cloud Compute Engine Web UI

Or with the command line interface

1. // CREATE INSTANCE WITH 4 vCPUs and 5 GB MEMORY
2. gcloud compute instances create my-vm --custom-cpu 4 --custom-memory 5

Customize for speed and

Runs on Google’s private
workload type (e.g. include
fiber network
GPUs and TPUs)
Optional Creating a VM on Google
Demo Cloud Platform
Processing and visualizing earthquake data on GCP
 Big Data and ML Products

Compute Power Storage Networking

Security
1.2 billion photos and
videos are uploaded to
Google Photos every day.

Total size of over 13 PB

of photo data.

1PB or 400 hours of video

uploaded every minute
Leverage Google’s 99.999999999% durability storage

Cloud
Storage
 Creating a Cloud Storage bucket for your data is easy

UI Cloud
Storage

gsutil mb -p [PROJECT_NAME] -c [STORAGE_CLASS]

CLI -l [BUCKET_LOCATION] gs://[BUCKET_NAME]/
Typical big data analytics workloads run in Regional Storage
Organization
Organization Cloud Storage buckets are one of
many resources of the Google
Cloud Platform
Team A Team B
Folders
You can collaborate with many
Product 1 Product 2
other teams in your organization
across many projects
Projects
Dev Test project Production

Resources
BigQuery Cloud Storage Compute Engine BigQuery
dataset bucket instance dataset
Got data? Quickly migrate your data to the cloud using gsutil tool
Google Cloud Platform Project

Bucket
Copy

Objects Cloud
Data and Storage
metadata

gsutil cp sales*.csv gs://acme-sales/data/

gsutil = google storage utility, cp = copy

 Big Data and ML Products

Compute Power Storage Networking

Security
Google’s private network carries as much as 40% of the
world’s internet traffic every day
 Google’s data center network speed enables
the separation of compute and storage

1 Petabit/sec of total bisection bandwidth

Servers doing compute Data can be “shuffled”

tasks don’t need to have the between compute workers
data on their disks at over 10GBs
 Google’s cable network spans the globe

FASTER (US, JP, TW) 2016

Havfrue (US,IE, DK) 2019

SJC (JP, HK, SG) 2013

HK-G (HK, GU) 2019

Unity (US, JP) 2010
Curie (CL, US) 2019
PLCN (HK, LA) 2019

Monet (US, BR) 2017

Google Network

Edge points of Junior (Rio, Santos) 2017

presence >100
Edge node locations Tannat (BR, UY, AR) 2017
>7500
Indigo (SG, ID, AU) 2019
 Big Data and ML Products

Compute Power Storage Networking

Security
On-premise → you manage all security layers
On-
Responsibility premises

Content

Access policies

Usage

Deployment

Web app security

Identity

Operations

Access and authentication

Network security

OS, data, and content

Audit logging

Network

Storage and encryption

Hardware
 Google Cloud Platform offers fully-managed services
On- IaaS PaaS Managed
Responsibility premises services

Content
You Manage Access policies

Usage

Deployment
Google Managed
Web app security

Identity

Operations

Access and authentication

Network security

OS, data, and content

Audit logging

Network

Storage and encryption

Hardware
Agenda
Google Cloud Platform infrastructure
● Compute
● Storage
● Networking
● Security
Big data and ML products
● Google innovation timeline
● Choosing the right approach
Lab: Exploring Public Datasets in BigQuery
Activity: Explore a customer use case
 Google invented new data processing methods as it grew

GFS MapReduce Bigtable Dremel Flume Spanner TensorFlow TPU

Colossus Megastore Millwheel

Pub/Sub F1

http://research.google.com/pubs/papers.html
 Google Cloud opens up that innovation and infrastructure to you

Cloud Dataproc Bigtable BigQuery Dataflow Dataflow ML Engine AutoML

Storage

Cloud
Cloud Datastore Pub/Sub
Spanner
Storage
The suite of big data products on Google Cloud Platform

Storage Ingestion Analytics Machine Learning Serving

Cloud Compute Kubernetes Cloud Data Studio

Storage BigQuery Cloud ML
Engine Engine TPU Dashboards/BI

Cloud Cloud Cloud Cloud Cloud Cloud Dialogflow

SQL Spanner TensorFlow
Dataflow Composer Dataproc AutoML

Cloud Cloud
Cloud Cloud Cloud
Datastore Bigtable
Functions ... App
Pub/Sub Datalab Engine
ML APIs
Agenda
Google Cloud Platform infrastructure
● Compute
● Storage
● Networking
● Security
Big data and ML products
● Google innovation timeline
● Choosing the right approach
Lab: Exploring Public Datasets in BigQuery
Activity: Explore a customer use case
BigQuery has over 130 Public Datasets to explore
Demo Query 2 billion lines of code in
less than 30 seconds
Github on BigQuery
Lab Exploring Public Datasets using
the BigQuery Web UI
● Open BigQuery
● Query a Public Dataset
● Create a custom table
● Load data into a new table
● Querying basics
 Open Qwiklabs
Open an incognito window
1 4 Launch the course from My Learning
(or private/anonymous window)

2 Go to events.qwiklabs.com

Sign In with existing account or Join with

3 new account (with email you used to
register for the bootcamp)
Don’t remember what email you
used to register with or don’t
have access to it?

https://goo.gl/xrVBpM
If so, use the link or QR code to
add your email address for
access to the course material.
View your lab

Note: You can access the course

PDFs under Lecture Notes

Launch the lab and start on Lab 1

 Labs will last for 40
minutes or so (and
materials can be
accessed for 2 years)

Tip: Track your

progress with Score
X/15

Pro tip: Use the table of

contents on the right to
quickly navigate
 Do not click
End Lab until you are
done with that lab

(note: each lab is

independent)
Do Ask Questions

(we have a talented

team of experts to
help!)
Agenda
Google Cloud Platform infrastructure
● Compute
● Storage
● Networking
● Security
Big data and ML products
● Google innovation timeline
● Choosing the right approach
What you can do with GCP
Activity: Explore a customer use case
GO-JEK brings goods and services to over
2 million families in 50 cities in Indonesia
GO-JEK’s footprint nationwide
GO-JEK manages 5 TB+ per day for analysis
GO-JEK soon faced data scale and latency challenges

“Most of the reports are Day +1, so

we couldn’t identify the problems
as soon as possible.”
GO-JEK migrated their data pipelines to GCP

● High performance scalability with

minimal operational maintenance
● More granular data with high
velocity and less latency (stream
processing)
● The ability to solve business
problems with real time data
insights
GO-JEK architecture review
 GO-JEK ride-share supply/demand use case

Business question

Which locations have mismatched supply and demand in real time.

Challenge

We ping every one of our drivers every 10 seconds, which means 6

million pings per minute and 8 billion pings per day. How do we stream
and report on such a volume of data?
 Autoscale streaming pipelines with Cloud Dataflow

Driver location ping Autoscaling

Dataflow autoscale based Automatically adjust the number of workers based on demand
on throughput data

Read Driver location fr...

Running
29 sec

Parse Json to booking

16,258 elements/s
56 sec

Current workers
Map Drive location to...
16,258 element/s Target Workers
3 sec
Visualize demand/supply mismatches with GIS data
 Your turn: Analyze real customer big data use cases
1. Navigate to
cloud.google.com/
customers/.

2. Filter Products &

Solutions for Big Data
Analytics.

3. Find an interesting
customer use case.

4. Identify the key

challenges, how they
were solved with the
cloud, and impact.