48
Krishna Gade Why everything is an A/B test at Pinterest

Why everything is an A/B Test at Pinterest

Embed Size (px)

Citation preview

Page 1: Why everything is an A/B Test at Pinterest

Krishna Gade

Why everything is an A/B test at Pinterest

Page 2: Why everything is an A/B Test at Pinterest

Confidential

2

Page 3: Why everything is an A/B Test at Pinterest

What is Pinterest?

A visual bookmarking tool

Discover an inspiring ideaSave it to a board

Go do it

Page 4: Why everything is an A/B Test at Pinterest

Pinterest is a data product

Page 5: Why everything is an A/B Test at Pinterest
Page 6: Why everything is an A/B Test at Pinterest

Data Flow

App

events

Kafka

SecorSkyline

Pinball

Redshift

Pinalytics

Features

Qubole (Hadoop)

Singer

Page 7: Why everything is an A/B Test at Pinterest

Confidential

Metrics

7

Page 8: Why everything is an A/B Test at Pinterest

Confidential

User Lifecycle

New Users

Active Users

Dormant Users

Acquisition

Activation

Churn

Churn Resurrection

Page 9: Why everything is an A/B Test at Pinterest

Confidential

Terminology

● MAU – Monthly Active User● WAU – Weekly Active User● DAU – Daily Active User● WAR – Weekly Active Repinner● WAC – Weekly Active Clickthrougher● WARC – Weekly Active Repinner or Clickthrougher● 1d7 – User who comes back at-least once in 7 days after

signup● 1rc7 - User who repins/clicks at-least once in 7 days

after signup

Page 10: Why everything is an A/B Test at Pinterest

Confidential

MAUs

● Current progress to quarterly goal● MAU forecast● MAUs by app/gender/country● MAU Accounting

Page 11: Why everything is an A/B Test at Pinterest

Confidential

MAU Accounting

Page 12: Why everything is an A/B Test at Pinterest

Confidential

Acquisition Metrics

● Total Signups● Signups by App● Signups by Referrer● Invites Sent● Unique Invite Senders● Invite Signups

Page 13: Why everything is an A/B Test at Pinterest

Confidential

Activation Metrics

Activation Rate = 1rc7/signups

● Activation Rate by App/Gender/Referrer● Overall signups to 1rc7● Signup engagement funnel● Signups to WAU 35 days after signup

Page 14: Why everything is an A/B Test at Pinterest

Confidential

Engagement Metrics● *AU ratios

● # of Emails sent● # of Emails opened● # of Emails clicked● # of Notifications sent● # of Notifications clicked

Page 15: Why everything is an A/B Test at Pinterest

Confidential

Resurrection Metrics

● # of Users who weredormant for 28+ daysbut came back, split byplatform, referrer etc

Page 16: Why everything is an A/B Test at Pinterest

Confidential

Takeaways

● Emphasize on activation (process of converting a new user to MAU)

● Activation critical to long-term growth● Deeply understand how different

segments of users interact

Page 17: Why everything is an A/B Test at Pinterest

Confidential

Automated Metrics

17

Page 18: Why everything is an A/B Test at Pinterest

Confidential

Page 19: Why everything is an A/B Test at Pinterest

Confidential

Chart creation

Page 20: Why everything is an A/B Test at Pinterest

Confidential

Reuse existing query result

Page 21: Why everything is an A/B Test at Pinterest

Confidential

Metrics creation

Page 22: Why everything is an A/B Test at Pinterest

Confidential

A/B Experimentation

22

Page 23: Why everything is an A/B Test at Pinterest

Confidential

Categories of Experiments

● A/B testing new features● Copy testing● SEO changes● Infrastructure updates

Page 24: Why everything is an A/B Test at Pinterest

Confidential

A/B Testing Framework

Page 25: Why everything is an A/B Test at Pinterest

Confidential

Page 26: Why everything is an A/B Test at Pinterest

Confidential

Page 27: Why everything is an A/B Test at Pinterest

Confidential

Page 28: Why everything is an A/B Test at Pinterest

Confidential

A/B Testing Dashboard

Page 29: Why everything is an A/B Test at Pinterest

Confidential

Page 30: Why everything is an A/B Test at Pinterest

Confidential

p-value (via chi-squared test)

Statistical tests for validation

Page 31: Why everything is an A/B Test at Pinterest

Confidential

Impact of Badging

Page 32: Why everything is an A/B Test at Pinterest

Confidential

Impact of Badging

● 7% lift in DAU● 2.5% lift over long term● Effective at increasingengagement levels

Page 33: Why everything is an A/B Test at Pinterest

Confidential

Copy Testing

Definition: Testing different variations of text we display to the user.

● A/B Testing copy was a tedious process● Even when we did this we were optimizing

only for English

Page 34: Why everything is an A/B Test at Pinterest

Confidential

New Approach

● Ability to run a copy test w/o changing a line of code.

● Support for translating text dynamically● Use Multi-armed bandits to converge to showing

best performing text

Page 35: Why everything is an A/B Test at Pinterest

Confidential

Multi Armed Bandits

Page 36: Why everything is an A/B Test at Pinterest

Confidential

SEO

36

50+ Billion Pinscategorized by people

into more than

1 Billion Boards

Page 37: Why everything is an A/B Test at Pinterest

Confidential

37

What makes SEO so challenging?

What is the Impact of the new SEO feature?

Will this new change affect SEO?

Uhhh...Dunno...

Uhhh...Ask Google.

Page 38: Why everything is an A/B Test at Pinterest

Confidential

KEEPCALM

and

RUN ANSEO EXPERIMENT

Page 39: Why everything is an A/B Test at Pinterest

Confidential

What is an SEO experiment?

39

Regular AB Experiment SEO Experiment

Metrics User behavior Search engine referred traffic

Unique id for hashing userid / sessionid page url

Page 40: Why everything is an A/B Test at Pinterest

Confidential

What is an SEO experiment?

Launch

● Measure the gap ● Before vs After launch

Page 41: Why everything is an A/B Test at Pinterest

Confidential

Dashboard

41

DIFF = enabled - control

Traffic Gain

Page 42: Why everything is an A/B Test at Pinterest

Confidential

42

Decision Making 1 2

3

4

SHIP!

SHIP!

SHUT DOWN!VALIDATE

Page 43: Why everything is an A/B Test at Pinterest

Confidential

43

Experiment Driven SEO

What is the Impact of the new SEO feature?

We got XX% of traffic growth from it! Can we use more resources to make bigger impact?

No, the experiment result is flat, so let’s ship it.

SEO experiment

SEO experiment

Is my UI change affecting SEO negatively?

Page 44: Why everything is an A/B Test at Pinterest

Confidential

Why you should be A/B testing even your infrastructure

Page 45: Why everything is an A/B Test at Pinterest

Confidential

Infrastructure Changes

● 2013: A new web framework● 2014: Pyapns● 2015: User service

Page 46: Why everything is an A/B Test at Pinterest

Confidential

Experiment results for replacing PyAPNs

Page 47: Why everything is an A/B Test at Pinterest

Confidential

Conclusions

● Figure out what success means● Define your metrics● Automate metric computation● A/B test everything!

Page 48: Why everything is an A/B Test at Pinterest

Confidential

Thank You!