New Metrics Engine to Help Drive UBER

t.uber.com/scala2016

Sasha OvsankinUBER

New Metrics Engine to Help Drive Uber

November 2016

1


“Transportation as reliable as running water, everywhere. for everyone”

2


Data & Analytics Engineer

About Me

Mathematical PhysicsLomonosov Moscow University

Contact

[email protected]://linkedin.com/in/sashaohttp://t.uber.com/scala2016

3

mailto:[email protected]


https://linkedin.com/in/sashao

https://linkedin.com/in/sashao

http://t.uber.com/scala2016



What Do We Work On

Fuel Uber’s innovation, make software release cycle more robust and

data driven

Experimentation Platform Uber Data Platform

Cutting-edge data platforms powering Uber’s intelligence

4


What This Talk Is About

Building a company-wide Metrics Platform is possible and practical,

and you should do it

5


Agenda

Why Metrics PlatformTechnologyProcessConclusion

6


How Do You Want Your Metrics?

Aligned

Reliable

Trusted

7


Uber Situation

● Over 450 cities in over 70 countries

● Lots of growth: ○ 1B rides by Dec 2015, 2B rides by

June 2016● Teams have high level of

independence

8


How do you make data-driven decisions in a business like that?

9


Metrics Platform = Technology + Process

10


Our Metrics PlatformArchitecture and Process

Engines

Registry

Council

Web-UI

Spark / Hive / Real Time

BI Tool UI

DS / Ops / Product

Definition DSL Query DSL

11


Our Metrics Platform

Easy & Powerful

Integrated

Lightweight Process

12


Metric Walkthrough

Metric hours active

English description

Hours spent by drivers logged-in and online in the driver app

SQL select sum(ac.minutes_active / 60) / count(*)from derived.driver_activity ac, right join dim.driver dr on dr.id=ac.driver_id

13


Metric walkthroughContinued

Add date select sum(ac.minutes_active / 60) / count(*)from derived.driver_activity ac, right join dim.driver dr on dr.id=ac.driver_idwhere ac.timestamp >= ‘2016-10-01’ and ac.timestamp < ‘2016-11-01’

In San Francisco select sum(ac.minutes_active / 60) / count(*)from derived.driver_activity ac, right join dim.driver dr on dr.id=ac.driver_idjoin dim.city c on c.id=ac.city_idwhere ac.timestamp >= ‘2016-10-01’ and ac.timestamp < ‘2016-11-01’ and c.name=’San Francisco’

14


Metric walkthroughContinued#2

Group by experiment treatment

select sum(ac.minutes_active / 60) / count(*)from derived.driver_activity ac, right join dim.driver dr on dr.id=ac.driver_idjoin dim.city c on c.id=ac.city_idjoin xp.user_experiment xp on xp.user_id=dr.idwhere ac.timestamp >= ‘2016-10-01’ and ac.timestamp < ‘2016-11-01’ and c.name=’San Francisco’and xp.experiment_key=’crm_driveronboarding_wcdrip’group by xp.treatment

Group by driver type

select sum(ac.minutes_active / 60) / count(*)from derived.driver_activity ac, right join dim.driver dr on dr.id=ac.driver_idjoin dim.city c on c.id=ac.city_idjoin xp.user_experiment xp on xp.user_id=dr.id join model.driver dm on dm.id=dr.idwhere ac.timestamp >= ‘2016-10-01’ and ac.timestamp < ‘2016-11-01’ and c.name=’San Francisco’and xp.experiment_key=’crm_driveronboarding_wcdrip’group by xp.treatment, model.driver.type

15


Complicated

Unmanageable

Fragile

16


Anatomy of a Metric

Preaggregationtransformations

Preaggregationtransformations

Aggregation

Aggregation

Post-aggregationformulasInput

Input

Input

Results

Dimensions

Metric definitions

Filters

FinalJoin

dim1

dim2

…

m1

m2

...

dim1

dim2

…

m1

m2

...

select sum(ac.minutes_active / 60) / count(*)from derived.driver_activity ac,right join dim.driver dr on dr.id=ac.driver_idwhere ac.timestamp >= ‘2016-10-01’ and ac.timestamp < ‘2016-11-01’group by driver.city

17


Metric = Formula + Query

select(core.avg_driver_hours_active)where(dim_city.name===”San Fransisco”)over(days(7) upto today)groupBy(driver_model.category, user_experiment.treatment)

select sum(ac.minutes_active / 60) / count(*)from derived.driver_activity ac,right join dim.driver dr on dr.id=ac.driver_idjoin dim.city c on c.id=ac.city_idwhere ac.timestamp >= ‘2016-10-01’ and ac.timestamp < ‘2016-11-01’and c.name=’San Francisco’...

avg_driver_hours_active = sum(agg.driver_activity.minutes_active / 60) / count(dim.driver)

FormulaQuery

18


The Metrics DSL: Formula

val hours_online = driver_activity.minutes_active / 60val all_drivers = count(dim_driver)val avg_driver_hours_online = sum(hours_online) / all_drivers

sum count

/

/

driver_activity.minutes_active 60

dim_driver

19


The Metrics DSL: Query

val query= select(avg_driver_hours_online) where(dimDriver.partner_city_id==="San Francisco") over(days * 7 towards today) groupBy(dimDriver.partner_city_id)

val df= engine.toDF(query)

DSL DataFrame Output20


✔ Easy & Powerful

Integrated

Lightweight Process

21


The Engine Core

Company Schema Repository

22


The Engine Core


Table Schemas

Foregn key Relationships

23


The Engine Core


Table Schemas

Foreign key Relationships

Engine Configuration

24


The Engine Core


Table Schemas



Engine Core

Query

25


The Engine Core


Table Schemas



Engine Core

Query

Execution Plan

26


✔ Easy & Powerful

✔ Integrated

Lightweight Process

27


Our Metrics PlatformArchitecture and Process

Engines

Registry

Council

Web-UI

Spark / Hive / Real Time

BI Tool UI

DS / Ops / Product

Definition DSL Query DSL

28


Metric Creation Process

29


Metric Management Web UI

Video link: https://youtu.be/we3q6O4eZIg 30

http://www.youtube.com/watch?v=we3q6O4eZIg

https://youtu.be/we3q6O4eZIg




Our Metrics PlatformTechnology

✔ Easy & Powerful

✔ Integrated

✔ Lightweight Process

31


● Experimentation● Product groups● Financial reporting● Real time decision making● Fraud detection

Users

32


● Futher adoption within Uber● Further work on DSL● More Engines● Real Time ● Open Source?

Future Direction

Interested?

http://t.uber.com/[email protected]

33






What this talk was about

Building company-wide Metrics Platform is possible and practical,

and you should do it

34


The Metrics Platform Team

Contact us:● http://t.uber.com/scala2016● [email protected]

35

We are hiring!






Questions?

36


Thank you

Proprietary and confidential © 2016 Uber Technologies, Inc. All rights reserved. No part of this document may be

reproduced or utilized in any form or by any means, electronic or mechanical, including photocopying, recording, or

by any information storage or retrieval systems, without permission in writing from Uber. This document is intended

only for the use of the individual or entity to whom it is addressed and contains information that is privileged,

confidential or otherwise exempt from disclosure under applicable law. All recipients of this document are notified

that the information contained herein includes proprietary and confidential information of Uber, and recipient may not

make use of, disseminate, or in any way disclose this document or any of the enclosed information to any person

other than employees of addressee to the extent necessary for consultations with authorized personnel of Uber.

Image credits: ● Erik bij de Vaate● Bernard Spragg

37

New Metrics Engine to Help Drive UBER

Software

Transcript of New Metrics Engine to Help Drive UBER