Más contenido relacionado Sifting, Sorting and Scanning for Gold – The Real-Time User Data Problem2. NoSQL in Digital Advertising
• Display advertising – today’s state of the art
Display advertising today s state of the art
• Real‐time data in digital advertising
• C
Case study
d
– User data store & cookie mapping
• Market differentiation
• Moving beyond the last click
g y
Citrusleaf Confidential © 2011 Citrusleaf. All rights reserved. 2
3. Display Advertising
• Lie:
– Content determines the value of an impression
• Lie:
– The value of an individual user is constant
– The decision funnel
The decision funnel
• Truth:
– The right ad to the right user at the right time
The right ad, to the right user, at the right time
© 2011 Citrusleaf. All rights reserved. 3
4. Today’s state of the art
• Impression logs => user segments (warehouse)
• Third part data => user segments
• Region segments
• Real‐time per‐user factor
• Frequency caps
• Session management for cookie less users
Session management for cookie‐less users
• Mapping external partner IDs
• Build optimization tables based on log analysis
Build optimization tables based on log analysis
Simple math determines highest value ad
Simple math determines highest value ad
© 2011 Citrusleaf. All rights reserved. 4
5. Real‐time data requirements
• Scalable and
flexible
fl ibl
• 100% availability
• Billion object
support
• High performance
with low hardware
cost
• Sophisticated data
eviction policies
© 2010 Citrusleaf. All rights reserved. 5
6. Case Study
Map store User store
Key
y cookie string or
g Key internal user id
partner id Value Segment data
Value internal user id (8 bytes) Frequency caps
Other optimization data
North America North America
1B ~ 1.5B objects 500M ~ 800M objects
Write load:
Write load: Write load:
Write load:
1k ~ 2k per second 10K ~ 20K per second
Read load: Read load:
10k ~ 50k per second
p 20k ~ 100k per second
Configuration Configuration
DRAM backed by disk
DRAM backed by disk 1T 4T user storage
1T ~ 4T user storage
Lowest latency (0.4 ms) Low latency (0.8 ms)
© 2011 Citrusleaf. All rights reserved. 6
7. Market differentiation
• Control over impressions
Control over impressions
• Quality and variety of inventory
y y y
• Better insight generation & tracking
• And:
– High quality user understanding
Trend: multiple user stores
© 2011 Citrusleaf. All rights reserved. 7
8. The next challenge
Weakness: “last ad takes all”
Performance is judged externally by the last
P f i j d d t ll b th l t
advertiser before a conversion
So… game the system to place strategically
New technology allows real‐time understanding of
user behavior
But …
Will advertisers stand for it?
© 2011 Citrusleaf. All rights reserved. 8
9. Indexed user behavior storage
• Store 100B+ user
behavior objects
b h i bj t
(1B per day, 90
days)
• Index by user_id,
giving sub‐second
access to entire
behavior chain
• O ti i d f
Optimized for
rotational media
• Time‐based eviction
Time‐based eviction
© 2010 Citrusleaf. All rights reserved. 9
10. Indexed user behavior storage
• Solves the “needle in a haystack” problem
• Allows immediate behavioral triggering
All i di b h i l i i
• Removes ETL and Map/Reduce scans
• Allows sophisticated attribution models,
(and advanced optimization and reporting)
The right ad,
g ,
to the right user,
at the right time
t th i ht ti
© 2011 Citrusleaf. All rights reserved. 10