Design Ads CTR Prediction Pipeline — System Design Practice | AmanAI Lab

Sheet›Design Ads CTR Prediction Pipeline

GoogleMetaAmazonMSFT

45:00

Hard

The Problem

Problem

Design the ads click-through rate (CTR) prediction system for a social media platform with 1 billion daily users and 5 million advertisers.

The system needs to:

•Predict the probability that a user clicks an ad (CTR)
•Serve predictions in ≤ 10ms per ad impression
•Handle 1 trillion impressions per day
•Retrain the model with fresh data every hour

What you'll be assessed on

Feature engineering for ads (user + ad + context), model architecture (Wide & Deep, DLRM), real-time vs batch feature serving, and the calibration of predicted probabilities for auction pricing.

Scale & Constraints

▸1B DAU, 5M advertisers
▸1 trillion impressions/day (~11M impressions/second)
▸CTR prediction latency ≤ 10ms
▸Hourly model retraining with last 7 days of data
▸Predicted probability must be well-calibrated (used in auction)
▸Handle high-cardinality categorical features (user IDs, ad IDs: billions)

Must Cover0/10

Hints (if stuck)

💡 Pre-compute and cache user and ad embeddings offline — dot product at serving is <1ms.

💡 In-batch negatives: treat other ads in the batch as negatives for the two-tower model.

💡 Wide part handles memorisation (feature crosses), deep part handles generalisation.

💡 Calibration is critical: uncalibrated probabilities lead to poor auction pricing and revenue loss.

💡 Feature freshness: user recent clicks must be within seconds; historical aggregates can be 1h stale.

0 words · auto-saved

Back to sheet

Mock interviewTake quizFlashcards

Problem

Design the ads click-through rate (CTR) prediction system for a social media platform with 1 billion daily users and 5 million advertisers.

The system needs to:

•Predict the probability that a user clicks an ad (CTR)

•Serve predictions in ≤ 10ms per ad impression

•Handle 1 trillion impressions per day

•Retrain the model with fresh data every hour

What you'll be assessed on