how spotify uses big data for fast product iterations | wouter de bie - spotify

Post on 14-Jun-2015

316 Views

Category:

Data & Analytics

1 Downloads

Preview:

Click to see full reader

DESCRIPTION

How Spotify Uses Big Data For Fast Product Iterations | Wouter de Bie - Spotify during Social Media Week Rotterdam 26-09-2014 In this talk Wouter de Bie looks at how Spotify uses data and Big Data technologies to make fast iterations on the Spotify product. Some of the questions he’ll try to answer are ”Why is fast product iteration important for us?”, ”How does data tie into this?” and ”What is it we do to achieve this”?

TRANSCRIPT

September 26, 2014

Wouter de Bie Big Data Architect @xinit / wouter@spotify.com

Big Data for fast product iterations to drive user growth

Big Data for fast product iterations to drive user

growth

September 26, 2014

Big Data for fast product iterations to drive user

growth

September 26, 2014

Big Data for fast product iterations to drive user

growth

September 26, 2014

Big Data •  40 million Monthly Active Users •  20+ million tracks •  57 countries •  2 TB of compressed data from users per day •  70 TB of data generated in Hadoop each day •  700 node Hadoop cluster •  28 PB of storage •  8500 – 12.000 jobs per day

September 26, 2014

Big Data for fast product iterations to drive user

growth

September 26, 2014

•  We were the first ones to do free streaming •  Not any longer..

•  We need to build the best product •  But nobody has done this before.. What does best mean?

Why fast iteration?

8

How do we develop our product?

Think it Build it Ship it Tweak it

September 26, 2014

Tweak it

Think it

Build it

Ship it

Understand consumers, behavior, opportunities, what to test

Build prototypes, ensure right metrics / tracking is in place

Understand what works, measure, evaluate, learn, optimize

Manage roll out, evaluate test groups, understand problems

Support with data and analytics

10

A/B testing •  Select a test group of (< all) users

•  Select a control group (e.g. all

remaining users)

•  Let the test group try a new version

of a feature

•  Gather metrics about the test and

control group

•  Compare the groups and roll out the

new feature if the test group

performs better

What did we find?

Ad frequency

Song   Song   Ads   Song   Song  Song   Song   Song   Song  

Song   Song   Ad   Song   Song  Song   Song   Song  Ad   Ad   Song  

VS

Result: no difference

Personalized email subjects

Result: 200% higher CTR

Radio

Result: Constantly improving radio experience

Sign-up flow

Signup button

Group B - test

Group A - control

Performance

Time

Download

Okay, listen to music

A - control 50% of users

B - test 50% of users

100% of users

Massive boost: ca 3x!

Layout of signup -button

Result: 3x more sign-ups!

“This signup-flow will kill it!”

•  Auto play music •  Great cover art •  Clear download instructions

•  Screenshot in the background

Result

Some history..

September 26, 2014

September 26, 2014

September 26, 2014

September 26, 2014

September 26, 2014

September 26, 2014

September 26, 2014

And not only for ourselves..

September 26, 2014

Jay-Z

Artist analytics

Artist analytics

Artist analytics

Some words of wisdom..

September 26, 2014

36

Only 10% will lead to a change – Google after 12.000 tests

It’s a numbers game

37

“80% of the times, we are wrong about what consumers want”

Leverage your data!

38 Section name

You’re the expert, but prepare for the truth

Thank you!

September 26, 2014

top related