1.
From Scratch
1
Joe Crobak
@joecrobak
!
Tuesday, June 24, 2014
Axium Lyceum - New York, NY
BUILDING A
DATA PIPELINE

2.
INTRODUCTION
2
Software Engineer @ Project Florida
!
Previously:
•Foursquare
•Adconion Media Group
•Joost

3.
OVERVIEW
3
Why do we care?
Defining Data Pipeline
Events
System Architecture

4.
4
DATA PIPELINES ARE EVERYWHERE

5.
RECOMMENDATIONS
5
http://blog.linkedin.com/2010/05/12/linkedin-pymk/

6.
RECOMMENDATIONS
6
Clicks
Views
Recommendations
http://blog.linkedin.com/2010/05/12/linkedin-pymk/

7.
AD NETWORKS
7

8.
AD NETWORKS
8
Clicks
Impressions
User Ad Profile

9.
SEARCH
9
http://lucene.apache.org/solr/

10.
SEARCH
10
Search Rankings
Page Rank
http://www.jevans.com/pubnetmap.html

11.
A / B TESTING
11
https://flic.kr/p/4ieVGa

12.
A / B TESTING
12
https://flic.kr/p/4ieVGa
A conversions
B conversions
Experiment Analysis

13.
DATA WAREHOUSING
13
http://gethue.com/hadoop-ui-hue-3-6-and-the-search-dashboards-are-out/

14.
DATA WAREHOUSING
14
http://gethue.com/hadoop-ui-hue-3-6-and-the-search-dashboards-are-out/
key metrics
user events
Data Warehouse

15.
15
WHAT IS A DATA PIPELINE?

16.
DATA PIPELINE
16
A Data Pipeline is a unified system for
capturing events for analysis and
building products.

17.
DATA PIPELINE
17
click data
user events
Data Warehouse
web visits
email sends
…
Product Features
Ad Hoc analysis
•Counting
•Machine Learning
•Extract Transform Load (ETL)

18.
DATA PIPELINE
18
A Data Pipeline is a unified system for
capturing events for analysis and
building products.

19.
19
EVENTS

20.
EVENTS
20
Each of these actions can be thought of as an
event.

21.
COARSE-GRAINED EVENTS
21
•Events are captured as a by-product.
•Stored in text logs used primarily for
debugging and secondarily for analysis.

22.
COARSE-GRAINED EVENTS
22
127.0.0.1 - - [17/Jun/2014:01:53:16 UTC] "GET / HTTP/1.1" 200 3969!
IP Address Timestamp Action Status
•Events are captured as a
•Stored in
debugging and secondarily for analysis.

23.
COARSE-GRAINED EVENTS
23
Implicit tracking—i.e. a “page load” event is a
proxy for ≥1 other event.
!
e.g. event GET /newsfeed corresponds to:
•App Load (but only if this is the first time
loaded this session)
•Timeline load, user is in “group A” of an A/B
Test
These implementations details have to be known at analysis time.

24.
FINE-GRAINED EVENTS
24
Record events like:
•app opened
•auto refresh
•user pull down refresh
!
Rather than:
•GET /newsfeed

25.
FINE-GRAINED EVENTS
25
Annotate events with contextual
information like:
•view the user was on
•which button was clicked

26.
FINE-GRAINED EVENTS
26
Decouple logging and analysis. Create events
for everything!

27.
FINE-GRAINED EVENTS
27
A couple of schema-less formats are popular
(e.g. JSON and CSV), but they have
drawbacks.
•harder to change schemas
•inefficient
•require writing parsers

28.
SCHEMA
28
Used to describe data, providing a contract
about fields and their types.
!
Two schemas are compatible if you can read
data written in schema 1 with schema 2.

29.
SCHEMA
29
Facilities automated analytics—summary
statistics, session/funnel analysis, a/b testing.

30.
SCHEMA
30
https://engineering.twitter.com/research/publication/the-unified-logging-infrastructure-for-data-analytics-at-twitter
Facilities automated analytics—summary
statistics, session/funnel analysis, a/b testing.

31.
SCHEMA
31
client:page:section:component:element:action e.g.:
!
iphone:home:mentions:tweet:button:click!
!
Count iPhone users clicking from home page:
!
iphone:home:*:*:*:click!
!
Count home clicks on buttons or avatars:
!
*:home:*:*:{button,avatar}:click

32.
32
KEY COMPONENTS

33.
EVENT FRAMEWORK
33
For easily generating events from your
applications

34.
EVENT FRAMEWORK
34
For
applications

35.
BIG MESSAGE BUS
35
•Horizontally scalable
•Redundant
•APIs / easy to integrate

36.
BIG MESSAGE BUS
36
•Scribe (Facebook)
•Apache Chukwa
•Apache Flume
•Apache Kafka*
!
•Horizontally scalable
•Redundant
•APIs / easy to integrate
* My recommendation

37.
DATA PERSISTENCE
37
For storing your events in files for batch
processing

38.
DATA PERSISTENCE
38
For
processing
Kite Software Development Kit
http://kitesdk.org/
!
Spring Hadoop
http://projects.spring.io/spring-hadoop/

39.
WORKFLOW MANAGEMENT
39
For coordinating the tasks in your data
pipeline

40.
WORKFLOW MANAGEMENT
40
… or your own system written
in your own language of choice.
*
For
pipeline

41.
SERIALIZATION FRAMEWORK
41
Used for converting an Event to bytes on
disk. Provides efficient, cross-language
framework for serializing/deserializing data.

42.
SERIALIZATION FRAMEWORK
42
•Apache Avro*
•Apache Thrift
•Protocol Buffers (google)
Used for
disk
framework for serializing/deserializing data.

43.
BATCH PROCESSING AND AD HOC
ANALYSIS
43
•Apache Hadoop (MapReduce)
•Apache Hive (or other SQL-on-Hadoop)
•Apache Spark

44.
SYSTEM OVERVIEW
44
Application
logging
framework
data
serialization
Message Bus
Persistant
Storage
Data
Warehouse
Ad hoc
Analysis
Product
data flow
workflow engine
Production
DB dumps

45.
SYSTEM OVERVIEW (OPINIONATED)
45
Application
logging
framework
data
serialization
Message Bus
Persistant
Storage
Data
Warehouse
Ad hoc
Analysis
Product
data flow
workflow engine
Production
DB dumps
Apache Avro
Apache Kafka
Luigi

46.
NEXT STEPS
46
This architecture opens up a lot of possibilities
•Near-real time computation—Apache
Storm, Apache Samza (incubating), Apache
Spark streaming.
•Sharing information between services
asynchronously—e.g. to augment user
profile information.
•Cross-datacenter replication
•Columnar storage

47.
LAMBDA ARCHITECTURE
47
Term coined by Nathan Marz (creator of
Apache Storm) for hybrid batch and real-
time processing.
!
Batch processing is treated as source of truth,
and real-time updates models/insights
between batches.

48.
LAMBDA ARCHITECTURE
48
http://lambda-architecture.net/

49.
SUMMARY
49
•Data Pipelines are everywhere.
•Useful to think of data as events.
•A unified data pipeline is very powerful.
•Plethora of open-source tools to build
data pipeline.

50.
FURTHER READING
50
The Unified Logging Infrastructure for Data
Analytics at Twitter
!
The Log: What every software engineer should
know about real-time data's unifying
abstraction (Jay Kreps, LinkedIn)
!
Big Data by Nathan Marz and James Warren
!
Implementing Microservice Architectures

51.
THANK YOU
51
Questions?
!
Shameless plug: www.hadoopweekly.com

52.
52
EXTRA SLIDES

53.
WHY KAFKA?
53
• https://kafka.apache.org/
documentation.html#design
• Pull model works well
• Easy to configure and deploy
• Good JVM support
• Well-integrated with the LinkedIn stack

54.
WHY LUIGI?
54
• Scripting language (you’ll end up writing
scripts anyway)
• Simplicity (low learning curve)
• Idempotency
• Easy to deploy

55.
WHY AVRO?
55
• Self-describing files
• Integrated with nearly everything in the
ecosystem
• CLI tools for dumping to JSON, CSV