Google·Machine Learning Engineer·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say

Jun 2026

Summary

Google system design interview for what seemed like an ML or ads-adjacent role. One meaty question about ad conversion prediction, which sounds straightforward until you actually have to spec it out end to end.

Questions Asked (1)

Design a system to predict YouTube ad conversions.

System DesignProduct Analytics & MetricsTechnical Trade-offs

Author's notes

I started with feature engineering and the interviewer kind of let me run with it, which in hindsight was a trap because I spent way too long there and barely touched the serving layer.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business objective and success metrics, then outline the end-to-end ML pipeline from data collection to deployment. Emphasize trade-offs between model complexity, latency, and scalability, and discuss how to handle challenges like delayed feedback and class imbalance.

Pro tip: Highlight the importance of defining a clear proxy for conversions (e.g., view-through conversions) and designing an experimentation framework to measure incremental lift, as correlation vs. causation is critical in ad systems.

1. Clarify Requirements and Metrics

Ask questions to understand the goal: predict conversions (e.g., purchase, sign-up) after ad exposure. Define success metrics like AUC, calibration, and business KPIs (e.g., ROI, incremental conversions).

2. Data and Feature Engineering

Identify data sources: user demographics, ad content, context, historical interactions. Discuss feature engineering for user, ad, and context, and handle challenges like delayed conversion labels and class imbalance.

3. Model Selection and Training

Choose models (e.g., logistic regression, GBDT, deep neural networks) based on scale and latency. Address techniques for imbalanced data (e.g., resampling, focal loss) and delayed feedback (e.g., survival analysis, positive-unlabeled learning).

4. Evaluation and Validation

Design offline evaluation (time-based split, holdout) and online A/B testing. Discuss metrics like AUC, PR-AUC, calibration, and business impact. Consider counterfactual evaluation to measure incremental lift.

5. Deployment and Monitoring

Plan for scalable serving (low latency, high throughput), model versioning, and monitoring for drift and performance degradation. Discuss retraining frequency and feedback loops.

Key Points to Mention

Handling delayed feedback: conversions may occur hours or days after ad exposure, requiring techniques like survival analysis or positive-unlabeled learning.
Class imbalance: conversions are rare events, so use appropriate sampling, weighting, or loss functions.
Feature engineering: incorporate user history, ad creative, context (time, device), and cross features.
Model choice trade-offs: balance accuracy, interpretability, latency, and scalability (e.g., deep models vs. GBDT).
Evaluation: use time-based validation, calibration, and online A/B tests to measure incremental lift.
Scalability: design for low-latency serving and handle high-volume data with distributed training.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.