Data Processing in RTB: Fast, Cheap and 98% Accurate
summary
Covers real-time analytics for RTB (real-time bidding) at RuTarget, which processes a billion ad-impression requests per day. Describes using streaming randomized algorithms and probabilistic data structures to cheaply and quickly estimate metrics such as the number of unique users with ~98% accuracy.
AI-written summary · about
chapters (12)
- 0:00RTB data processing intro
- 1:10Real-time ad auction challenges
- 2:39Probabilistic sampling for accuracy
- 4:09Reservoir sampling explained
- 10:17Outliers in data streams
- 12:23HyperLogLog for unique counts
- 14:42Parallel processing with sketches
- 15:03Count-Min sketch for frequencies
- 23:12Q&A
- 25:24Memory vs accuracy trade-offs
- 26:30Sketches in production
- 28:45Building histograms from streams
related talks
16mThrough the looking glass – Following the rabbit down the data… Whole
53mReal-Time Big Data Analytics with R
29mAI and ML for Computational Advertising
19mSession: Re-Engagement In 2026: How To Design & Scale Winning Retargeting Campaigns
26mWin More Effectively in Real-Time Bidding
23mStream Algorithms for Big Data Processing Tasks
08mReal-Time Data Systems: Speed, Scale, and Resilience
01h 01mHow to Compute CTR for 100M Items in Real Time and Survive