際際滷

際際滷Share a Scribd company logo

Reel bid insightd-eproject

M
Mudit Uppal

1. Real-time bidding (RTB) is a programmatic advertising method where automated ad buying and selling occurs within milliseconds of an ad impression. 2. The document discusses Reelbid, a demand-side platform that handles RTB. It proposes using sampling and hierarchical techniques from a research paper to more efficiently select bids and determine pricing to improve return on investment. 3. The goal is to design an architecture to handle millions of queries per second for real-time bidding while efficiently scaling without adding nodes to reduce costs.

1 of 23
Download to read offline
Reelbid Real time ad bidding DSP ( programmatic) MUDIT UPPAL INSIGHT DATA ENGINEERING, SV
What is Real Time Bidding(RTB)? Image source:
Reel bid insightd-eproject
Demand side platform - Reelbid Mashable.com (publisher with inventory) Sell Side Platform OpenRTB server Auction service Ad exchange OpenRTB client Bidder service Data management platform Data broker Some advertiser
Motivation Really interesting & new* data engineering challenges Current State of the art Googles open-bidder still in alpha(not for public) In 2014, according to Business Insider Intelligence, ad revenue topped $15 billion. Real-time bidding, and in particular, mobile and video real-time bidding, lead the way for that growth. Business Insider Intelligence estimates say RTB revenue will pass $26 billion by the end of 2020, which far surpasses the $11.7 billion from this year. ( *Disclaimer: NO background in Ad tech ) Latency ( 100 ms ) Queries Per Second (~2 - 3 million bids/second) + Infrequent bids
Some Sample AdExchange Reelbid 1. Hey do you want to bid on this users impression?: 2. Yes! Heres my bid: 100 ms "id" : "32a69c6ba388f110487f9d1e63f77b22d86e916b", "banner": { "h": 250, "w": 300, "name": "mashable.com", "domain": "http://www.example.com", .. "page": "http://easy.example.com/easy?cu=13824; cre=mu;target=_blank", "ref" : "http://refer+url", "publisher": { "id": "qqwer1234xgfd", "name": "site_name", "domain": "my.site.com" "device": { "ua": "Mozilla/5.0(KHTML, like Gecko) Version/5.1.7 Safari/534. 57.2", "ip": "192.168.5.5", "geo": { "lat": 37.789, "lon": -122.394, "country": "USA", "city": "San Francisco", "region": "CA", "zip" : "94105", "type": 2 } "user": { "buyeruid" : "89776897686798fwe87rtryt8976fsd7869678", "id": "55816b39711f9b5acf3b90e313ed29e51665623f", "gender": "M", "yob": 1975, "customdata": "Data-asdfdwerewr", "data": [{ "id": "pub-demographics", "name": "data_name", "segment": [{ "id" : "345qw245wfrtgwertrt56765wert", "name" : "segment_name", "value": "segment_value } { "id": "eb85349d-03c3-44f4-a77b- 824f7221d116", "seatbid": [{ "bid": [{ "id": "bid1", "impid": "eb85349d-03c3-44f4-a77b- 824f7221d116", "price": 0.1, "adm": "<div>Ad Creative</div>", "adomain": [ "http://www.example.com/clickthrough"
Goals - Create a more efficient bidding framework using sampling and hierarchical techniques (as described in the paper) - Ways to Scale without adding nodes and save costs
Demo
Reel bid insightd-eproject
A different approach to RTB A paper published by in IEEE for Data Mining-- (source of truth) Instead of comparing bid requests with user profile<U,P> we look at bid requests only. Compensate computational resources with price (and not accept and bid on everything) ROI with Gaussian/Gamma models are much better
Architecture 1.0
Architecture 2.0 11 2 3
How do we handle 2-3 million QPS? (efficient scaling)
Sampling! Ziggurat Algorithm for sampling random values within milli seconds (also most memory efficient )
How do we decide the bid price?
Y < f(x) Cost(Budget) Resources 1. Looking at overall compute resources 2. User input filtering
Algorithm Request selection Based on Utilization Selection policy based on reward models Filter 1 Filter 2 Filter 3
Collecting server system metrics - Graphite - Carbon | StatsD | Collectd - Redis
Cost + ROI
Testing - Event-sim - Parallec io (ebay) [8K async HTTP requests per second] - Smaato AdX
Challenges - To process that file in 80 ms - Throttling issues - Networking challenges - Programming/debugging/testing in a distributed environment is a very time- consuming task
About me Surf, Football and micro-controllers [ Building products ] Data Scientist, Planetary/Fusion Network Ltd (NY) MS Comp Science, Media, Business Mgmt - University of California, Santa Barbara MUDIT UPPAL
Learning at Insight: Learning by doing Before Insight: Batch processing, Data Science Analytics After Insight: Stream processing [ Spark streaming, Storm, Kafka, Cassandra ] + learning from feedback, the right tool for the right job, More debugging/testing in a distributed computing with polyglot approach + Thinking asynchronously

More Related Content

Reel bid insightd-eproject

  • 1. Reelbid Real time ad bidding DSP ( programmatic) MUDIT UPPAL INSIGHT DATA ENGINEERING, SV
  • 2. What is Real Time Bidding(RTB)? Image source:
  • 4. Demand side platform - Reelbid Mashable.com (publisher with inventory) Sell Side Platform OpenRTB server Auction service Ad exchange OpenRTB client Bidder service Data management platform Data broker Some advertiser
  • 5. Motivation Really interesting & new* data engineering challenges Current State of the art Googles open-bidder still in alpha(not for public) In 2014, according to Business Insider Intelligence, ad revenue topped $15 billion. Real-time bidding, and in particular, mobile and video real-time bidding, lead the way for that growth. Business Insider Intelligence estimates say RTB revenue will pass $26 billion by the end of 2020, which far surpasses the $11.7 billion from this year. ( *Disclaimer: NO background in Ad tech ) Latency ( 100 ms ) Queries Per Second (~2 - 3 million bids/second) + Infrequent bids
  • 6. Some Sample AdExchange Reelbid 1. Hey do you want to bid on this users impression?: 2. Yes! Heres my bid: 100 ms "id" : "32a69c6ba388f110487f9d1e63f77b22d86e916b", "banner": { "h": 250, "w": 300, "name": "mashable.com", "domain": "http://www.example.com", .. "page": "http://easy.example.com/easy?cu=13824; cre=mu;target=_blank", "ref" : "http://refer+url", "publisher": { "id": "qqwer1234xgfd", "name": "site_name", "domain": "my.site.com" "device": { "ua": "Mozilla/5.0(KHTML, like Gecko) Version/5.1.7 Safari/534. 57.2", "ip": "192.168.5.5", "geo": { "lat": 37.789, "lon": -122.394, "country": "USA", "city": "San Francisco", "region": "CA", "zip" : "94105", "type": 2 } "user": { "buyeruid" : "89776897686798fwe87rtryt8976fsd7869678", "id": "55816b39711f9b5acf3b90e313ed29e51665623f", "gender": "M", "yob": 1975, "customdata": "Data-asdfdwerewr", "data": [{ "id": "pub-demographics", "name": "data_name", "segment": [{ "id" : "345qw245wfrtgwertrt56765wert", "name" : "segment_name", "value": "segment_value } { "id": "eb85349d-03c3-44f4-a77b- 824f7221d116", "seatbid": [{ "bid": [{ "id": "bid1", "impid": "eb85349d-03c3-44f4-a77b- 824f7221d116", "price": 0.1, "adm": "<div>Ad Creative</div>", "adomain": [ "http://www.example.com/clickthrough"
  • 7. Goals - Create a more efficient bidding framework using sampling and hierarchical techniques (as described in the paper) - Ways to Scale without adding nodes and save costs
  • 10. A different approach to RTB A paper published by in IEEE for Data Mining-- (source of truth) Instead of comparing bid requests with user profile<U,P> we look at bid requests only. Compensate computational resources with price (and not accept and bid on everything) ROI with Gaussian/Gamma models are much better
  • 13. How do we handle 2-3 million QPS? (efficient scaling)
  • 14. Sampling! Ziggurat Algorithm for sampling random values within milli seconds (also most memory efficient )
  • 15. How do we decide the bid price?
  • 16. Y < f(x) Cost(Budget) Resources 1. Looking at overall compute resources 2. User input filtering
  • 17. Algorithm Request selection Based on Utilization Selection policy based on reward models Filter 1 Filter 2 Filter 3
  • 18. Collecting server system metrics - Graphite - Carbon | StatsD | Collectd - Redis
  • 20. Testing - Event-sim - Parallec io (ebay) [8K async HTTP requests per second] - Smaato AdX
  • 21. Challenges - To process that file in 80 ms - Throttling issues - Networking challenges - Programming/debugging/testing in a distributed environment is a very time- consuming task
  • 22. About me Surf, Football and micro-controllers [ Building products ] Data Scientist, Planetary/Fusion Network Ltd (NY) MS Comp Science, Media, Business Mgmt - University of California, Santa Barbara MUDIT UPPAL
  • 23. Learning at Insight: Learning by doing Before Insight: Batch processing, Data Science Analytics After Insight: Stream processing [ Spark streaming, Storm, Kafka, Cassandra ] + learning from feedback, the right tool for the right job, More debugging/testing in a distributed computing with polyglot approach + Thinking asynchronously