M9-002M9: Senior ML & AI System DesignRecSys, Ads & Search ArchitectureHard
⚡Flashcard Drill➔Mastery:

Design a Real-Time Ads Ranking, eCPM Calibration & Budget Pacing Architecture

🎯 Active Recall Self-Check (Draft First, Verify After)
Login-free practice tracker: recall core answer, audit pitfalls, and log review status.
Step 1 · Draft your concise answer & assumptions
Step 2 · Check Pitfalls & Follow-up Inquiries
Next: M9-003 →
📐 Mathematical Definition
eCPM=Bid⋅p^CTR⋅1000,logit(p)=logit(q)+ln⁡r\text{eCPM} = \text{Bid} \cdot \hat{p}_{\text{CTR}} \cdot 1000,\quad \text{logit}(p)=\text{logit}(q)+\ln r
⚡ One-Sentence Core Answer
Enforce auction integrity via well-calibrated eCPM bids, negative-downsampling log-odds correction, delayed conversion survival reweighting, and closed-loop PID throttling.

📌 Key Takeaways

  • •
    L1 Monetary Invariant: ranking AUC isn't enough—miscalibrated pCTR corrupts GSP auction bids
  • •
    L2 Sampling Calibration: shift log-odds by +ln(neg_sample_rate) and verify ECE < 1%
  • •
    L2 Delayed Feedback: reweight early unconverted clicks via survival/FNW importance weights
  • •
    L3 Pacing Control: PID probabilistic throttling smooths budget burn cleanly across hours

📐 Mathematical Derivations

True log-odds follow logit(p)=logit(q)+ln⁡r\text{logit}(p)=\text{logit}(q)+\ln r, while 10-second PID feedback adjusts campaign participation rates smoothly.

🏭 Production Trade-offs

Probabilistic throttling preserves truthful auction valuation while shedding redundant fine-ranking load for ahead-of-schedule campaigns.
⚠️ Common Interview Pitfalls
  • ✕
    Shipping negative-downsampled models without ln(r) calibration bias
  • ✕
    Marking immature recent clicks as hard negatives before conversion windows mature
🎯 Interviewer Follow-up Inquiries
  • ?
    Why is ECE critical in ads auctions while organic feeds care mainly about AUC?
  • ?
    How do you damp PID pacing oscillations on low-budget campaigns?
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.
🎯 Drill This Card Now➔
📚

Associated Knowledge Base Guides & Mindmaps

← PreviousM9-001: Design a Billion-Scale Four-Stage Feed Recommendation Architecture📋 Back to BankNext →M9-003: Design a Cold-Start Item & Creator Exploration Pipeline